求解器系统

dispatchatlas.solve 拥有求解器元数据、选择、调度构造、修复、性能保障、基线族、可选适配器，以及 NDSO 族。

该包仅导入 dispatchatlas.core。基准烟雾集成测试位于 tests/solve/，因此运行时包不依赖于具体的基准生成器。

可运行示例： examples/compare_solvers.py 并行调度一组求解器并以一个已证明的精确最优解为其锚定。

注册表

SolverRegistry 存储带有丰富元数据的无状态求解器工厂：

能力标签，如 capacity-aware、precedence-aware、repair、local-search，以及 ndso
支持的目标，如 makespan、energy，以及 cost
通过能力标签表达的已声明约束
默认停止准则
确定性或种子化随机的重放行为
解编码（permutation、mapping、assignment，或 native）
可选依赖声明与商业后端披露
用于分阶段导出的证据-层级可见性
一个规范引用，或一个显式的引用-不适用理由

每个具名求解器族携带一个失败关闭的引用：一个具有规范开创性来源却省略其引用的族无法被构造，而一个没有单一规范来源的族记录原因而非捏造一个。

from dispatchatlas.solve import SolverCapability, default_solver_registry
 
registry = default_solver_registry()
metaheuristics = registry.select(
    required_capabilities=(SolverCapability.METAHEURISTIC,),
    objective="makespan",
)

注册表目录

注册表中的每个求解器，呈现于一个可排序、可搜索的表格中。该表及其总数由 default_solver_registry() 生成，因此下面的计数可从行本身计算。表格上方的一张能力覆盖图汇总有多少个注册表求解器声明每一项已声明能力。

Generated from the solver registry: 87 solvers across 6 groups — constructive (5), dispatching (14), exact (7), learning (11), metaheuristic (47), ndso (3).

How many of the 87 registry solvers advertise each declared capability — a metadata-only view of the registry’s breadth (a constructive baseline through to many-objective and native solvers). Hover or focus a bar to read its count.

Search solversGroupGroup by

Showing 87 of 87 solvers.

Solver registry — 87 rows, build-inlined from the public registry bundle.
		Supported objectives	Notes
`adpso`	metaheuristic	makespan, energy, cost	Inspect caveats success-rate feedback is noisy on small swarms competitor adpso currency anchor Sensors (MDPI), 2022 encoding adapter random-key evidence class runnable-baseline mechanism swarm search with success-adaptive descending inertia strengths inertia tracks search progress instead of a fixed ramp venue tier indexed-peer-reviewed
`age-moea-ii`	metaheuristic	makespan, lateness, fairness, energy, cost, carbon	Inspect caveats geometry estimate is noisy on tiny first fronts competitor age-moea-ii evidence class runnable-baseline geometry Newton-Raphson curvature fit of the sum(x^p)=1 surface mechanism adaptive front-geometry estimation with geodesic survival strengths spread faithful to convex, linear, and concave fronts
`ant-colony`	metaheuristic	makespan, energy, cost	Inspect mechanism rate-gated adjacent swaps of the incumbent order metaheuristic ant-colony realization no pheromone trail matrix or cooperating colony is implemented; the citation marks lineage, not mechanism fidelity scheduling contract permutation-decode-to-schedule
`apparent-tardiness-cost`	dispatching	makespan, lateness, energy, cost	Inspect dispatch rule apparent-tardiness-cost priority basis highest apparent-tardiness-cost index first (unweighted, k=2)
`arithmetic-optimization`	metaheuristic	makespan, energy, cost	Inspect caveats every member is regenerated around the best vector each iteration, so the returned order is the best observed rather than a monotone incumbent discrete adaptation Abualigah, L., Diabat, A., & Abd Elaziz, M. (2021). Intelligent workflow scheduling for Big Data applications in IoT cloud computing environments. Cluster Computing, 24, 2957-2976. doi:10.1007/s10586-021-03291-7 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search whose accelerator schedule shifts from wide-scatter division and multiplication moves to tight addition and subtraction moves around the best vector as the budget elapses strengths the two operator pairs give a clean exploration-to-exploitation handover governed only by the iteration ratio
`artificial-bee-colony`	metaheuristic	makespan, energy, cost	Inspect caveats the per-component neighbor move explores slowly on long priority vectors, so many iterations may be needed at large task counts discrete adaptation Pan, Q.-K., Tasgetiren, M. F., Suganthan, P. N., & Chua, T. J. (2011). A discrete artificial bee colony algorithm for the lot-streaming flow shop scheduling problem. Information Sciences, 181(12), 2455-2468. doi:10.1016/j.ins.2009.12.025 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search with three phases per iteration: a greedy neighbor perturbation of every member toward a random peer, a fitness-weighted resample that perturbs promising members again, and a re-initialization of any member that fails to improve for a set number of trials strengths the fitness-weighted resample concentrates effort on promising members while the trial-limit re-initialization sustains diversity
`artificial-fish-swarm`	metaheuristic	makespan, energy, cost	Inspect caveats each behaviour re-scores candidate positions, so an iteration costs several schedule evaluations per member discrete adaptation Ibadi, A. M., & Rahman, R. A. (2024). Modified artificial fish swarm algorithm to solve unrelated parallel machine scheduling problem under fuzzy environment. AIMS Mathematics, 9(1), 1679-1705. doi:10.3934/math.20241679 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search where each member adopts the better of a swarm move toward nearby members' centre and a follow move toward the best nearby member, each falling back to a local random probe when its move condition fails strengths the crowding test spreads members across several basins rather than collapsing them onto one centre
`beam-search`	constructive	makespan, energy, cost	Inspect beam width 3 mechanism breadth-bounded constructive search priority basis best topological completion score
`branch-and-bound`	exact	makespan, energy, cost	Inspect commercial false dependency ortools dependency extra exact dependency purpose branch-and-bound / branch-and-cut adapter dependency requirement optional exact method branch-and-bound license note OR-Tools is Apache-2.0 licensed; no commercial terms. optimality verification solves-to-optimal-verified
`ccgp`	metaheuristic	makespan, energy, cost	Inspect caveats rule evaluation cost grows with population and graph size competitor ccgp currency anchor Computers & Operations Research (Elsevier), 2024 evidence class runnable-baseline mechanism cooperative coevolution of priority-rule trees strengths learns instance-feature priority rules instead of orders venue tier top-tier-operations-research
`clpso`	metaheuristic	makespan, energy, cost	Inspect decode rule random-key encoded solver clpso encoding adapter random-key repair policy topological-normalize transfer identity
`cma-es`	metaheuristic	makespan, energy, cost	Inspect caveats the random-key encoding is a documented adaptation of the continuous method to the discrete order space encoding tasks sequenced by ascending random key, repaired to precedence feasibility evidence class runnable-baseline mechanism covariance-matrix-adaptation evolution strategy over a random-key encoding of the task order recombination weighted mean of the best mu offspring with rank-one and rank-mu covariance updates step size control cumulative step-size adaptation with the h_sigma stall guard strengths adapts an anisotropic search distribution to the objective landscape without hand-tuned scaling
`cpop`	dispatching	makespan, energy, cost	Inspect mechanism combined upward-and-downward rank list scheduling priority basis upward plus downward rank; critical tasks first rank basis combined-rank realization rank-ordered dispatch; resource and mode binding by the serial constructor's earliest-finish rule, a documented variant of the published critical-path-processor pinning
`critical-path-tabu`	metaheuristic	makespan, energy, cost	Inspect caveats path relinking between elite solutions is not realized; a documented reduction of the published method competitor critical-path-tabu evidence class runnable-baseline mechanism critical-path block-boundary tabu search relates to tabu-search: the mechanism-diversified competitor realization strengths moves concentrate on the makespan-defining chain
`cuckoo-search`	metaheuristic	makespan, energy, cost	Inspect caveats the Levy move is greedy per member, so progress depends on the incumbent and the step scale rather than a monotone population discrete adaptation Marichelvam, M. K., Prabaharan, T., & Yang, X.-S. (2014). Improved cuckoo search algorithm for hybrid flow shop scheduling problems to minimize makespan. Applied Soft Computing, 19, 93-101. doi:10.1016/j.asoc.2014.02.005 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search that steps every member toward the best-so-far solution by a heavy-tailed Levy displacement, keeps the step only when it improves, and re-initializes a fixed fraction of the worst members each iteration strengths the heavy-tailed step occasionally makes a large jump that escapes local optima while the discovery step sustains diversity
`d-clpso`	metaheuristic	makespan, energy, cost	Inspect decode rule spv encoded solver d-clpso encoding adapter spv repair policy topological-normalize search domain [-1.0, 1.0] transfer identity
`d-depso`	metaheuristic	makespan, energy, cost	Inspect hybrid of differential-evolution + particle-swarm scheduling contract permutation-decode-to-schedule scheduling variant d-depso
`d-lshade`	metaheuristic	makespan, energy, cost	Inspect decode rule rounding encoded solver d-lshade encoding adapter rounding repair policy topological-normalize transfer identity
`dan-dual-attention`	learning	makespan, energy, cost	Inspect documented adaptation faithful dual-attention architecture -- the operation message attention block over the precedence window and the machine message attention block over the machine-competition graph -- adapting the published amortized PPO-clip training to single-instance REINFORCE over seeded rollouts, greedy order decode, and the platform's uniform per-machine durations that collapse the (operation, machine) action to the operation order the serial decoder scores evidence class runnable-baseline learned paradigm policy-gradient dual-attention construction mechanism dual-attention network + policy-gradient PDR scheduler optional extra learning policy basis dual-attention network with an operation message attention block over the precedence window and a machine message attention block over the machine-competition graph, scoring compatible operation-machine pairs for a size-agnostic dispatching policy training single-instance REINFORCE over seeded rollouts with a terminal normalized-improvement reward and a mean-return baseline
`decima-dag-rl`	learning	makespan, energy, cost	Inspect documented adaptation faithful Decima two-level DAG message passing and REINFORCE policy gradient; the composite (node, parallelism-limit) action collapses to node selection because the platform exposes no parallelism-limit control, per-node critical-path features replace the reference executor and job-level features that have no platform counterpart, and single-instance training with a greedy decode and a terminal normalized schedule-score reward adapt the published cross-instance training and time-in-system reward to the platform's per-instance solve contract evidence class runnable-baseline learned paradigm policy-gradient GNN construction (REINFORCE) mechanism Decima two-level DAG-GNN + REINFORCE construction scheduler optional extra learning policy basis two-level DAG message passing with per-node, per-DAG, and global summaries training single-instance REINFORCE over seeded rollouts with a running-mean baseline and a terminal normalized-improvement reward
`decision-diagram-sequencing`	exact	makespan, energy, cost	Inspect commercial false exact method decision-diagram license note Native pure-Python decision-diagram branch and bound; no third-party backend and no licensing constraints. native task cap 12 optimality verification solves-to-optimal-verified
`differential-evolution`	metaheuristic	makespan, energy, cost	Inspect mechanism positional nudge toward a uniformly drawn donor order metaheuristic differential-evolution realization a discrete recombination, not the cited difference-vector arithmetic; the citation marks lineage, not mechanism fidelity scheduling contract permutation-decode-to-schedule
`earliest-deadline`	dispatching	makespan, energy, cost	Inspect dispatch rule earliest-deadline priority basis earliest declared deadline first (EDD / deadline-aware)
`earliest-finish-time`	dispatching	makespan, energy, cost	Inspect dispatch rule earliest-finish-time priority basis earliest reachable finish first (release + duration)
`earliest-start`	dispatching	makespan, energy, cost	Inspect dispatch rule input-order priority basis input order
`epso`	metaheuristic	makespan, energy, cost	Inspect caveats path gathering can over-constrain loosely coupled graphs competitor epso currency anchor Electronics (MDPI), 2023 encoding adapter random-key evidence class runnable-baseline mechanism workload-biased-initialization swarm with path gathering strengths quality-seeded swarm over a reduced particle dimension venue tier indexed-peer-reviewed
`exhaustive-enumeration`	exact	makespan, energy, cost	Inspect commercial false exact method exhaustive-enumeration license note Native pure-Python exact enumeration; no third-party backend and no licensing constraints. native task cap 8 optimality verification solves-to-optimal-verified
`firefly-algorithm`	metaheuristic	makespan, energy, cost	Inspect caveats the pairwise attraction is quadratic in the population size, and positions move unconditionally so the returned order is the best observed rather than a monotone incumbent discrete adaptation Sayadi, M. K., Hafezalkotob, A., & Naini, S. G. J. (2013). Firefly-inspired algorithm for discrete optimization problems: An application to manufacturing cell formation. Journal of Manufacturing Systems, 32(1), 78-84. doi:10.1016/j.jmsy.2012.06.004 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search attracting every priority vector toward every brighter member by a distance-decaying step plus a randomization term that shrinks over the budget strengths distance-decaying attraction lets subgroups converge on several basins before the randomization term fades
`fjsp-hgnn-drl`	learning	makespan, energy, cost	Inspect documented adaptation faithful HGNN architecture; single-instance training, greedy order decode, and uniform per-machine durations adapt the published amortized FJSP scheduler to the platform's per-instance task-order model evidence class runnable-baseline learned paradigm policy-gradient GNN construction (PPO) mechanism heterogeneous-GNN + PPO construction scheduler optional extra learning policy basis heterogeneous graph attention over operations and machines training single-instance proximal policy optimization over seeded rollouts with a terminal normalized-improvement reward
`genetic-algorithm`	metaheuristic	makespan, energy, cost	Inspect mechanism elite-pair order crossover with adjacent-swap mutation metaheuristic genetic-algorithm realization selection-crossover-mutation loop over a permutation population; faithful in kind to the cited scheme at baseline scale scheduling contract permutation-decode-to-schedule
`grasshopper-optimization`	metaheuristic	makespan, energy, cost	Inspect caveats the pairwise social forces are quadratic in the population size, and positions move unconditionally each iteration; on a Largest Order Value decode the published update adds little beyond its initial population, because it replaces each position with the target plus the force sum rather than perturbing it, and the resulting displacement peaks near a tenth of the gap between adjacent priorities -- short of the separation a rank swap needs. The shortfall is scale-free in the priority span and does not close with population size, so it is a property of the method on this encoding rather than a tuning choice discrete adaptation Naveen Durai, S. K., & Duraisamy, B. (2022). Hybrid Invasive Weed Improved Grasshopper Optimization Algorithm for Cloud Load Balancing. Intelligent Automation & Soft Computing, 34(1), 467-483. doi:10.32604/iasc.2022.026020 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search displacing every member by the sum of pairwise attraction/repulsion forces from all others, scaled toward the best-so-far target by a comfort coefficient that decays over the budget strengths the decaying comfort coefficient shrinks the repulsion zone so the swarm settles from exploration onto the target
`gravitational-search`	metaheuristic	makespan, energy, cost	Inspect caveats the pairwise forces are quadratic in the population size, and the velocity carries momentum that can overshoot on a coarse budget discrete adaptation Choudhary, A., Gupta, I., Singh, V., & Jana, P. K. (2018). A GSA based hybrid algorithm for bi-objective workflow scheduling in cloud computing. Future Generation Computer Systems, 83, 14-26. doi:10.1016/j.future.2018.01.005 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search that gives every member a mass from its fitness, accelerates each member by the summed pairwise attractive forces of the others (each force scaling with the product of masses and decaying with separation), and integrates a velocity and position under a coefficient that shrinks over the budget strengths heavier better-scoring members exert more pull, so the population is drawn toward high-quality regions as the coefficient decay sharpens the search
`greedy-completion`	dispatching	makespan, energy, cost	Inspect dispatch rule greedy-completion priority basis state-aware earliest completion next
`grey-wolf-optimizer`	metaheuristic	makespan, energy, cost	Inspect caveats positions move unconditionally each iteration, so the returned order is the best observed rather than a monotone incumbent discrete adaptation Hosseini Shirvani, M. (2022). A Novel Discrete Grey Wolf Optimizer for Scientific Workflow Scheduling in Heterogeneous Cloud Computing Platforms. Scientia Iranica. doi:10.24200/sci.2022.57262.5144 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search drawing every priority vector toward the three best-ranked members with an encircling coefficient that descends linearly over the budget strengths the descending coefficient shifts the search from exploration to exploitation with no per-instance tuning
`guided-local-search`	metaheuristic	makespan, energy, cost	Inspect caveats the (position, task) feature and duration cost are a documented adaptation of the paper's routing-edge feature model evidence class runnable-baseline feature a (position, task) placement whose cost is the task duration mechanism penalty-guided local search that augments the objective with penalties on overused placement features neighborhood best-improvement pairwise swaps on the augmented objective penalty update the maximum-utility placement (cost over one-plus-penalty) is penalized at each local optimum strengths escapes local optima by discouraging the features a stuck optimum keeps reusing, without restarts
`gurobi-exact`	exact	makespan, energy, cost	Inspect commercial true dependency gurobipy dependency extra exact-commercial dependency purpose commercial mixed-integer programming wrapper dependency requirement optional exact method milp license note Gurobi requires a commercial license; a free academic license is available. Never bundled. optimality verification solves-to-optimal-verified
`harris-hawks-optimization`	metaheuristic	makespan, energy, cost	Inspect caveats the besiege moves replace positions unconditionally, so the returned order is the best observed rather than a monotone incumbent discrete adaptation Ali, A., Shah, S. A. A., Al Shloul, T., Assam, M., Ghadi, Y. Y., Lim, S., & Zia, A. (2024). Multiobjective Harris Hawks Optimization-Based Task Scheduling in Cloud-Fog Computing. IEEE Internet of Things Journal. doi:10.1109/JIOT.2024.3391024 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search whose decaying escaping-energy value switches between an exploration phase and four besiege moves around the best member, two of which compare a direct move against a Levy-flight move strengths the energy schedule interleaves exploration and several exploitation intensities without per-instance tuning
`heft`	dispatching	makespan, energy, cost	Inspect mechanism upward-rank list scheduling priority basis nonincreasing upward rank rank basis upward-rank realization rank-ordered dispatch; resource and mode binding by the serial constructor's earliest-finish rule, the non-insertion variant of the published insertion-based earliest-finish-time selection
`ibea`	metaheuristic	makespan, lateness, fairness, energy, cost, carbon	Inspect caveats the pairwise indicator matrix is quadratic in the pool competitor ibea discrete adaptation Frutos, M., Mendez, M., Tohme, F., & Broz, D. (2013). Comparison of multiobjective evolutionary algorithms for operations scheduling under machine availability constraints. The Scientific World Journal, 2013, 418396. doi:10.1155/2013/418396 evidence class runnable-baseline indicator additive epsilon indicator on normalized objectives mechanism binary quality-indicator fitness with the additive epsilon indicator; the least-fit member is removed and the survivors are incrementally credited with its indicator contribution strengths a single tunable indicator drives convergence without reference points, weights, or nondominated sorting
`iterated-greedy-rs`	metaheuristic	makespan, energy, cost	Inspect acceptance fixed-temperature probabilistic acceptance of non-improving sequences caveats acceptance pressure tracks the problem's duration scale evidence class runnable-baseline mechanism destruction-reconstruction search from a best-insertion seed relates to iterated-greedy: the mechanism-diversified competitor realization strengths the canonical published iterated-greedy configuration for permutation scheduling temperature basis temperature_factor times mean task duration over ten, a documented adaptation of the published constant to the scalar-duration task model
`jaya-algorithm`	metaheuristic	makespan, energy, cost	Inspect caveats greedy replacement keeps only improving updates, so progress tracks the population's best and worst spread discrete adaptation Mishra, A. K., & Shrivastava, D. (2020). A discrete Jaya algorithm for permutation flow-shop scheduling problem. International Journal of Industrial Engineering Computations, 11(3), 415-428. doi:10.5267/j.ijiec.2019.12.001 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism parameter-less population search moving each priority vector toward the best and away from the worst member strengths carries no algorithm-specific parameters beyond population size and iteration budget
`l2d-disjunctive-gnn`	learning	makespan, energy, cost	Inspect documented adaptation faithful Graph Isomorphism Network over the disjunctive graph under the adding-arc strategy with the scheduled-indicator and completion-lower-bound raw features; single-instance training, a one-shot initial-state embedding with a greedy order decode instead of the reference per-step re-embedding, and a terminal normalized schedule-score reward adapt the published amortized size-agnostic dispatcher to the platform's per-instance task-order model evidence class runnable-baseline learned paradigm policy-gradient GNN construction (PPO) mechanism GIN disjunctive-graph + PPO dispatching rule optional extra learning policy basis graph-isomorphism network over the disjunctive graph with an average-pooled size-agnostic graph embedding training single-instance proximal policy optimization over seeded rollouts with a terminal normalized-improvement reward
`l2s-improvement`	learning	makespan, energy, cost	Inspect documented adaptation faithful DRL-guided improvement heuristic that starts from a complete feasible schedule, embeds its disjunctive graph with the topological and context-aware modules over the (p, est, lst) node features, and learns N5 critical-block moves by REINFORCE with the incumbent-improvement reward; the N5 neighborhood is realized over the platform's precedence-feasible task order as makespan-binding adjacent same-machine transpositions, and single-instance REINFORCE over a bounded improvement horizon adapts the published amortized training to the platform's per-instance task-order model evidence class runnable-baseline learned paradigm policy-gradient GNN improvement (REINFORCE) mechanism DRL-guided N5 improvement heuristic optional extra learning policy basis graph-neural network with a topological (GIN) module and a context-aware dual-context module over the complete-solution disjunctive graph, scoring N5 critical-block local-search moves training single-instance REINFORCE over the improvement rollout with the incumbent-improvement reward, keeping the best order found
`learned-priority-policy`	learning	makespan, energy, cost	Inspect learned tier disclosure two distinct learned representatives -- this policy-gradient priority rule and the value-based tabular Q-learning agent (rl-dispatching); no post-Decima DRL scheduler has consensus-baseline adoption mechanism policy-gradient learned priority dispatching rule policy basis linear priority over task features training REINFORCE over seeded precedence-feasible rollouts
`logic-based-benders-decomposition`	exact	makespan, energy, cost	Inspect commercial false dependency pulp dependency extra exact dependency purpose logic-based Benders assignment-master ILP adapter dependency requirement optional exact method logic-based-benders license note PuLP is BSD-licensed and ships the CBC open backend; the assignment master is solved through it and the per-facility scheduling subproblem is native pure Python. optimality verification solves-to-optimal-verified
`longest-processing-time`	dispatching	makespan, energy, cost	Inspect dispatch rule longest-processing-time priority basis longest duration first
`lshade`	metaheuristic	makespan, energy, cost	Inspect decode rule random-key encoded solver lshade encoding adapter random-key repair policy topological-normalize transfer identity
`marine-predators`	metaheuristic	makespan, energy, cost	Inspect caveats each iteration costs two schedule evaluations per member (the forage move and the FADs escape) discrete adaptation Chen, D., & Zhang, Y. (2023). Diversity-Aware Marine Predators Algorithm for Task Scheduling in Cloud Computing. Entropy, 25(2), 285. doi:10.3390/e25020285 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search that forages toward the best member by a Brownian phase early, a mixed Brownian/Levy phase in the middle, and a Levy phase late, escaping local optima with a fish-aggregating-device perturbation and keeping only improving moves strengths the phased Brownian-to-Levy schedule plus memory saving gives a monotone incumbent with a strong late-run local search
`matheuristic-restricted-neighbourhood`	metaheuristic	makespan	Inspect activity selection network-dependent: rank by the progressive/regressive-level body rule and select 50% of the activities for intense search envelope classical RCPSP only: finish-start precedence with no lag, renewable resources, single minimise-makespan objective; fails closed on generalized precedence, non-makespan or maximize objectives, and moldable, gang, or partial-execution tasks evidence class runnable-baseline exact neighbourhood the selected subproblem is solved by the native immediate-selection branch-and-bound (Brucker et al. 1998) over the pinned sub-instance future work the paper's higher-performance decomposition genetic-algorithm variant (Section 3.2.3) is not implemented; the faithful greedy-search form is realised instead mechanism greedy-search matheuristic: random-key sampling generates start schedules and an exact branch-and-bound re-optimises a selected activity body as a restricted neighbourhood search never worse the neighbourhood search accepts only improvements over the start schedule, so the result never regresses below the best sampled start schedule; near-optimal, not exact restriction full forward: the non-selected activities are pinned at their current start times as a latest-start bound and the selected subproblem is searched exactly without further truncation
`max-min`	dispatching	makespan, energy, cost	Inspect degenerate when tasks pin all resources through demands mechanism ready-set batch mapping by largest best completion priority basis largest minimum completion next realization the precedence-ready set stands in for the published independent-task batch; candidate and committed bindings follow the serial constructor's earliest-finish rule
`min-min`	dispatching	makespan, energy, cost	Inspect degenerate when tasks pin all resources through demands mechanism ready-set batch mapping by smallest best completion priority basis earliest minimum completion next realization the precedence-ready set stands in for the published independent-task batch; candidate and committed bindings follow the serial constructor's earliest-finish rule
`minimum-slack`	dispatching	makespan, energy, cost	Inspect dispatch rule minimum-slack priority basis least slack first
`moead`	metaheuristic	makespan, lateness, fairness, energy, cost	Inspect caveats weight-grid resolution bounds attainable spread evidence class runnable-baseline mechanism weighted-decomposition subproblems with neighborhood replacement strengths decomposition-based counterweight to dominance-based selection
`monte-carlo-tree-search`	metaheuristic	makespan, energy, cost	Inspect caveats the running best/worst reward normalization is a documented adaptation of UCB1's bounded-reward assumption evidence class runnable-baseline mechanism upper-confidence-bounds-for-trees search that builds a schedule by appending precedence-ready tasks reward the constructed order's score is normalized to [0, 1] against the running best and worst scores for UCB1 rollout partial orders are completed by appending uniformly random precedence-ready tasks selection UCB1 balances the best mean reward against the least-visited child by the exploration constant strengths concentrates simulations on the construction prefixes that have yielded the best schedules
`moth-flame-optimization`	metaheuristic	makespan, energy, cost	Inspect caveats the flame memory guarantees monotone best positions, but each member's spiral is unconditional discrete adaptation Abd Elaziz, M., Abualigah, L., & Attiya, I. (2021). Advanced optimization technique for scheduling IoT tasks in cloud-fog computing environments. Future Generation Computer Systems, 124, 142-154. doi:10.1016/j.future.2021.05.026 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search spiralling every member around one of the best positions in a sorted memory whose size shrinks over the budget, keeping the best positions after each spiral strengths the shrinking flame memory retains several elite positions early and converges onto the best late, with no tuning
`ndso-core`	ndso	makespan, energy, cost	Inspect anytime convergence false coupled rate temperature true decayed cell reinjection true encoding natively-discrete method variant ndso-core peer guidance true positional preference model true preference weighted plurality true source selection schedule true three source guidance true update magnitude schedule true
`ndso-fast`	ndso	makespan, energy, cost	Inspect anytime convergence true coupled rate temperature true decayed cell reinjection false encoding natively-discrete method variant ndso-fast peer guidance true positional preference model true preference weighted plurality false source selection schedule true three source guidance false update magnitude schedule true
`ndso-summit`	ndso	makespan, energy, cost	Inspect coordination inter-swarm-council cross swarm confidence weighted voting true encoding natively-discrete method variant ndso-summit overarching elite synthesis true
`neh`	constructive	makespan, energy, cost	Inspect mechanism nonincreasing-duration best-insertion construction priority basis nonincreasing-duration insertion
`nsga2`	metaheuristic	makespan, lateness, fairness, energy, cost	Inspect caveats diversity pressure weakens beyond three objectives evidence class runnable-baseline mechanism nondominated sorting with crowding-distance elitism strengths the default bi/tri-objective baseline in scheduling comparisons
`nsga3`	metaheuristic	makespan, lateness, fairness, energy, cost, carbon	Inspect caveats reference-point count grows with objective dimension competitor nsga3 evidence class runnable-baseline mechanism reference-point niching over nondominated fronts reference point design Das-Dennis structured simplex strengths diversity preservation in many-objective spaces
`ortools-cp-sat`	exact	makespan, energy, cost	Inspect commercial false dependency ortools dependency extra exact dependency purpose exact constraint-programming adapter dependency requirement optional exact method cp-sat license note OR-Tools is Apache-2.0 licensed; no commercial terms. optimality verification solves-to-optimal-verified
`particle-swarm`	metaheuristic	makespan, energy, cost	Inspect mechanism single-task attraction toward the global-best order metaheuristic particle-swarm realization no velocity state or personal-best memory is implemented; the citation marks lineage, not mechanism fidelity scheduling contract permutation-decode-to-schedule
`peft`	dispatching	makespan, energy, cost	Inspect mechanism optimistic-cost-table rank list scheduling priority basis nonincreasing mean optimistic cost rank basis oct-rank realization rank-ordered dispatch; resource and mode binding by the serial constructor's earliest-finish rule, a documented variant of the published O_EFT selection
`pulp-milp`	exact	makespan, energy, cost	Inspect commercial false dependency pulp dependency extra exact dependency purpose mixed-integer linear programming adapter dependency requirement optional exact method milp license note PuLP is BSD-licensed and ships the CBC open backend. optimality verification solves-to-optimal-verified
`residual-scheduling`	learning	makespan, energy, cost	Inspect documented adaptation faithful residual (shrinking, re-embedded) HGIN over unfinished operations; single-instance training, greedy order decode, terminal-objective reward with a running-mean baseline, hidden width reduced 256->128, and uniform per-machine durations adapt the published amortized job-shop scheduler to the platform's per-instance task-order model evidence class runnable-baseline learned paradigm policy-gradient GNN construction (REINFORCE, residual state) mechanism residual heterogeneous-GIN + REINFORCE construction scheduler optional extra learning policy basis residual heterogeneous-GIN over unfinished operations training single-instance REINFORCE over seeded rollouts with a terminal normalized-improvement reward and a running-mean baseline
`rl-dispatching`	learning	makespan, energy, cost	Inspect evidence class runnable-baseline learned paradigm value-based (tabular Q-learning) mechanism tabular Q-learning over dispatching-rule actions policy basis tabular action-value over dispatch rules relates to learned-priority-policy: the policy-gradient learned representative; this is the value-based Q-learning representative state abstraction coarse (scheduling-stage, ready-set-breadth) discretization, a documented adaptation of the published shop-state features training one-step Q-learning with a terminal normalized-improvement reward over seeded epsilon-greedy episodes
`rollout`	constructive	makespan, energy, cost	Inspect base policy shortest-processing-time static-priority completion budget a wall-clock budget degrades the lookahead to the bare base-policy completion of the committed prefix caveats cost grows with the base-policy completions scored, quadratic in the task count evidence class runnable-baseline guarantee the sequentially-consistent base policy makes the rollout order score no worse than the base policy it improves mechanism one-step base-policy lookahead sequential construction relates to shortest-processing-time: the base policy this rollout improves strengths recovers order quality a one-pass dispatch rule leaves on the table by scoring a full completion per ready candidate
`rvea`	metaheuristic	makespan, lateness, fairness, energy, cost, carbon	Inspect caveats reference-vector count grows with objective dimension competitor rvea evidence class runnable-baseline mechanism angle-penalized-distance selection over reference vectors reference vector design unit-normalized Das-Dennis simplex strengths convergence-diversity balance that sharpens over the run
`salp-swarm`	metaheuristic	makespan, energy, cost	Inspect caveats positions move unconditionally each iteration, so the returned order is the best observed rather than a monotone incumbent discrete adaptation Jain, R., & Sharma, N. (2022). A Deadline-Constrained Time-Cost-Effective Salp Swarm Algorithm for Resource Optimization in Cloud Computing. International Journal of Applied Metaheuristic Computing, 13(1). doi:10.4018/IJAMC.292509 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous chain search whose leader is drawn toward the best-so-far food source by a coefficient that decays over the budget, and whose followers each move to the mean of their own and their predecessor's position strengths the follower averaging smooths the chain toward the leader while the decaying coefficient sharpens the leader's search
`sarsa-dispatching`	learning	makespan, energy, cost	Inspect discrete adaptation Aissani, N., Beldjilali, B., & Trentesaux, D. (2009). Dynamic scheduling of maintenance tasks in the petroleum industry: a reinforcement approach. Engineering Applications of Artificial Intelligence, 22(7), 1089-1103. doi:10.1016/j.engappai.2009.01.014; Zhang, Z., Zheng, L., Hou, F., & Li, N. (2011). Semiconductor final test scheduling with Sarsa(lambda, k) algorithm. European Journal of Operational Research, 215(2), 446-458. doi:10.1016/j.ejor.2011.05.052 evidence class runnable-baseline learned paradigm value-based on-policy TD control (SARSA) mechanism tabular on-policy SARSA over dispatching-rule actions policy basis tabular action-value over dispatch rules relates to rl-dispatching: the off-policy Q-learning representative; this is the on-policy SARSA representative training on-policy SARSA that bootstraps toward the next selected action-value, with a terminal normalized-improvement reward over seeded epsilon-greedy episodes
`scatter-search`	metaheuristic	makespan, energy, cost	Inspect caveats path relinking is precedence-repaired over the order space, a documented adaptation of the published combination method combination path relinking between the two reference orders of each subset competitor scatter-search evidence class runnable-baseline mechanism reference-set path-relinking search over precedence-safe orders realizes the elite-solution path relinking left unrealized in critical-path-tabu reference set a best-scoring quality subset plus a max-min positional-distance diversity subset strengths systematic recombination of diverse high-quality task orders
`selection-hyper-heuristic`	metaheuristic	makespan, energy, cost	Inspect acceptance late-acceptance hill-climbing against the cost recorded a fixed number of steps earlier evidence class runnable-baseline low level heuristics adjacent-swap, insertion-shift, relink-to-incumbent, bounded-local-search mechanism two-layer selection hyper-heuristic: adaptive heuristic selection over a fixed low-level heuristic set with a separate move-acceptance criterion relates to distinct from ccgp, a generation hyper-heuristic that evolves priority-rule trees rather than selecting among fixed low-level heuristics scheduling contract permutation-decode-to-schedule selection reinforcement-style adaptive credit; improving moves reward the applied heuristic, non-improving moves penalize it, with roulette-wheel proportional selection
`serial-sgs-justification`	constructive	makespan, energy, cost	Inspect caveats single-pass constructor; no search beyond the justification re-decodes; on unit-capacity (or demand-equals-capacity) problems the cumulative decode reduces to the scalar exclusive timeline citation doi pending verified publisher landing link recorded as the uri; the DOI digits are unverified and await operator confirmation evidence class runnable-baseline mechanism latest-finish-time serial construction with double justification, decoded through the cumulative capacity-profile placement so a resource packs concurrent activities up to its capacity realization double justification realized as finish- and start-sorted re-decodes through the serial constructor, a documented order-space adaptation of the published schedule-space shifts strengths justification improves serial-decoder schedules at the cost of only two extra decodes; the cumulative decode honors the capacity-aware capability on resources carrying demand below capacity
`shifting-bottleneck`	constructive	makespan, energy, cost	Inspect caveats heads and tails are longest-path bounds treating dependency lags as finish-to-start offsets; the serial constructor honors the exact timing evidence class runnable-baseline machine identity a distinct resource-demand set; demand-free tasks share a pool machine mechanism machine decomposition sequencing the bottleneck machine first reoptimization committed machines re-sequenced while later commitments tighten their release dates and tails strengths critical-path decomposition that resolves the most constraining resource before the rest subproblem single-machine release-date relaxation solved by Schrage's rule
`shortest-processing-time`	dispatching	makespan, energy, cost	Inspect dispatch rule shortest-processing-time priority basis shortest duration first
`simulated-annealing`	metaheuristic	makespan, energy, cost	Inspect mechanism adjacent-swap neighborhood under Metropolis acceptance metaheuristic simulated-annealing realization single-chain anneal, temperature cooling linearly to a 0.01 floor; faithful in kind to the cited scheme scheduling contract permutation-decode-to-schedule
`sine-cosine-algorithm`	metaheuristic	makespan, energy, cost	Inspect caveats positions move unconditionally each iteration, so the returned order is the best observed rather than a monotone incumbent discrete adaptation Jouhari, H., Lei, D., Al-qaness, M. A. A., Abd Elaziz, M., Ewees, A. A., & Farouk, O. (2019). Sine-Cosine Algorithm to Enhance Simulated Annealing for Unrelated Parallel Machine Scheduling with Setup Times. Mathematics, 7(11), 1120. doi:10.3390/math7111120 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search nudging every priority vector toward the best-so-far destination by a sine- or cosine-weighted step whose amplitude descends linearly over the budget strengths the oscillating step spans a wide region early and collapses onto the destination late with no per-instance tuning
`slim-self-labeling`	learning	makespan, energy, cost	Inspect documented adaptation faithful Pointer-Network-style encoder-decoder generative model trained by the self-labeling cross-entropy over the best sampled pseudo-label with neither an MDP nor reinforcement learning; the fifteen graph-node and eleven job-context features collapse to the platform-derivable subset, and single-instance self-labeling over a bounded step budget with a greedy decode adapts the published dataset-scale amortized training to the platform's per-instance task-order model evidence class runnable-baseline learned paradigm self-supervised generative construction mechanism pointer-network generative model + self-labeling optional extra learning policy basis pointer-network-style encoder-decoder generative model over the disjunctive graph with a multi-head-attention memory network training single-instance self-labeling: sample solutions, adopt the minimum-makespan sample as a pseudo-label, and minimize the cross-entropy toward it, using no reward or Markov Decision Process
`sms-emoa`	metaheuristic	makespan, lateness, fairness, energy, cost, carbon	Inspect caveats hypervolume cost grows with the objective count competitor sms-emoa evidence class runnable-baseline indicator exact dominated hypervolume via objective slicing mechanism dominated-hypervolume (S-metric) survival selection strengths reference-free spread and convergence balance
`spea2`	metaheuristic	makespan, lateness, fairness, energy, cost, carbon	Inspect caveats the pairwise distance matrix is quadratic in the pool competitor spea2 density k-th nearest neighbour with k = floor(sqrt N) discrete adaptation Wei, Feng, Tan, and Hagiwara. Flexible job shop scheduling multi-objective optimization based on an improved strength Pareto evolutionary algorithm. Advanced Materials Research, 186, 546-551, 2011. doi:10.4028/www.scientific.net/AMR.186.546 evidence class runnable-baseline mechanism strength-and-raw-fitness ranking with a k-th nearest-neighbour density estimator and distance-based archive truncation strengths the density estimator preserves a well-spread archive without reference points or scalarizing weights
`squeaky-wheel`	metaheuristic	makespan, energy, cost	Inspect analyzer blame equals the task completion time read from the constructed schedule assignments caveats the completion-time blame is a documented adaptation of the paper's domain-specific trouble measure evidence class runnable-baseline initialization duration-normalized prior with a seeded uniform perturbation of the starting priority ordering mechanism construct/analyze/prioritize search over a per-task priority vector that induces the construction sequence prioritizer blamed tasks' priorities rise by bump_strength times their makespan-normalized blame, pulling late tasks earlier strengths reprioritizes around the tasks that finish latest rather than searching the order space blindly
`sufferage`	dispatching	makespan, energy, cost	Inspect degenerate when tasks pin all resources through demands mechanism ready-set batch mapping by completion-loss gap priority basis largest second-best-minus-best completion gap next realization the precedence-ready set stands in for the published independent-task batch; candidate and committed bindings follow the serial constructor's earliest-finish rule
`surrogate-assisted-gp-hh`	learning	makespan, energy, cost	Inspect documented adaptation single static instance -> derived related tasks (makespan, lateness-weighted, flow-weighted scorings); 1-NN phenotypic-characterisation surrogate over ready-set reference situations; bounded population and generations for single-instance latency evidence class runnable-baseline extends ccgp learned paradigm genetic-programming hyper-heuristic (surrogate multitask) mechanism surrogate-assisted multitask GP hyper-heuristic policy basis evolved priority-rule trees decoded by precedence-safe list scheduling relates to ccgp: the cooperative-coevolution GP hyper-heuristic this surrogate multitask method extends training each generation a per-task 1-NN phenotypic-characterisation surrogate screens the merged offspring pool; only the reallocated survivors are really evaluated
`tabu-mrcpsp-mode-search`	metaheuristic	makespan, energy, cost	Inspect adaptation operates on the renewable-only, finish-start projection: the paper's generalized precedence relations (max lags, SF/FF), non-renewable and doubly constrained resources, and its inefficient-mode preprocessing are outside the core model and are projected away, not fabricated aspiration asymmetric: global and regional aspiration by objective override tabu status; influence always tilts move selection evidence class runnable-baseline inner scheduler a fixed mode assignment is projected onto the problem (each activity pinned to one mode) and decoded by the shared serial constructor, standing in for the paper's truncated RCPSP-GPR branch-and-bound long term memory frequency-based diversification and intensification phase switching after an initial data-collection phase mechanism tabu search over the mode-assignment neighborhood: each move reassigns one activity to another execution mode neighborhood single-activity re-mode -- all assignments differing from the incumbent in exactly one activity's mode tabu list dynamic random tenure in [sqrt(n), 3*sqrt(n)] on the reversing (activity, mode) attribute
`teaching-learning-optimization`	metaheuristic	makespan, energy, cost	Inspect caveats greedy replacement keeps only improving updates, so progress tracks the population's best member and mean discrete adaptation Xie, Z., Zhang, C., Shao, X., Lin, W., & Zhu, H. (2014). An effective hybrid teaching-learning-based optimization algorithm for permutation flow shop scheduling problem. Advances in Engineering Software, 77, 35-47. doi:10.1016/j.advengsoft.2014.07.006 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism parameter-less two-phase population search: a teacher phase draws every learner toward the best member and away from the population mean, and a learner phase draws it toward or away from a random peer by which is better strengths carries no algorithm-specific parameters beyond population size and iteration budget
`whale-optimization`	metaheuristic	makespan, energy, cost	Inspect caveats positions move unconditionally each iteration, so the returned order is the best observed rather than a monotone incumbent discrete adaptation Zhang, S., & Gu, X. (2023). A discrete whale optimization algorithm for the no-wait flow shop scheduling problem. Measurement and Control. doi:10.1177/00202940231180622 encoding continuous priority vector decoded to a permutation by the Largest Order Value rule, then precedence-repaired evidence class runnable-baseline mechanism continuous population search choosing per iteration a spiral move around the best member, an encircling move toward it, or a search move toward a random member, under a control value that descends linearly over the budget strengths the spiral and search moves preserve exploration while the descending control value sharpens exploitation, with no per-instance tuning

All objective support is declared registry metadata, not a performance claim. The solver recommender ranks these solvers by declared fit.

求解器适用性

哪些求解器适用于哪个基准族，读取自适用性矩阵。每个单元格都是声明的适用性，绝非性能主张：已验证的单元格引用具名的公开活动、引文或测试，而近似单元格则由声明的求解器能力和基准族特征推导得出。下表由公开适用性捆绑包生成，因此其状态总数可从各行自身计数。

Generated from the applicability matrix: 69 benchmark families by 87 public solvers. 10 verified · 5835 approximate · 158 not applicable.

Cell status is declared applicability, never a performance claim. Verified cells cite a named public campaign artifact, citation source, or test; approximate cells are derived from declared solver capabilities and benchmark-family traits and are labeled as such.

SearchBenchmark familySolverStatus

Showing 6003 of 6003 cells.

Benchmark-family by solver applicability — 6003 matching rows, showing page 1 of 121; build-inlined from the public applicability bundle.
			Basis
`accelerator-coscheduling`distributed-computing	`adpso`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; searches a continuous space decoded via the random-key-adapter with topological repair.
`accelerator-coscheduling`distributed-computing	`age-moea-ii`	Approximate	family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`ant-colony`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`apparent-tardiness-cost`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`arithmetic-optimization`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`artificial-bee-colony`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`artificial-fish-swarm`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`beam-search`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`branch-and-bound`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`ccgp`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`clpso`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; searches a continuous space decoded via the random-key-adapter with topological repair.
`accelerator-coscheduling`distributed-computing	`cma-es`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`cpop`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`critical-path-tabu`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`cuckoo-search`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`d-clpso`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; searches a continuous space decoded via the spv-adapter with topological repair.
`accelerator-coscheduling`distributed-computing	`d-depso`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`d-lshade`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; searches a continuous space decoded via the rounding-adapter with topological repair.
`accelerator-coscheduling`distributed-computing	`dan-dual-attention`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`decima-dag-rl`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`decision-diagram-sequencing`	Not applicable	native exact search caps at 12 tasks; family instances carry 96; optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`differential-evolution`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`earliest-deadline`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`earliest-finish-time`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`earliest-start`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`epso`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; searches a continuous space decoded via the random-key-adapter with topological repair.
`accelerator-coscheduling`distributed-computing	`exhaustive-enumeration`	Not applicable	native exact search caps at 8 tasks; family instances carry 96; optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`firefly-algorithm`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`fjsp-hgnn-drl`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`genetic-algorithm`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`grasshopper-optimization`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`gravitational-search`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`greedy-completion`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`grey-wolf-optimizer`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`guided-local-search`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`gurobi-exact`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`harris-hawks-optimization`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`heft`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`ibea`	Approximate	family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`iterated-greedy-rs`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`jaya-algorithm`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`l2d-disjunctive-gnn`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`l2s-improvement`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`learned-priority-policy`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`logic-based-benders-decomposition`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`longest-processing-time`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.
`accelerator-coscheduling`distributed-computing	`lshade`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; searches a continuous space decoded via the random-key-adapter with topological repair.
`accelerator-coscheduling`distributed-computing	`marine-predators`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling.
`accelerator-coscheduling`distributed-computing	`matheuristic-restricted-neighbourhood`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity; produces a static schedule for a family with dynamic arrivals/rescheduling; no repair or dispatching mechanism for the family's declared uncertainty.
`accelerator-coscheduling`distributed-computing	`max-min`	Approximate	optimizes one objective axis of a multi-objective family; family declares constraint features with no capability-tag counterpart: affinity.

Cell status is declared applicability, never a performance claim. Verified cells cite a named public campaign, citation, or test; approximate cells are derived from declared solver capabilities and benchmark-family traits. The solver recommender scores solvers against this matrix by scheduling family.

派遣基线

捆绑的确定性基线是构造式列表-调度求解器。每个控制任务顺序；串行构造器尊重每个任务声明的资源需求，并将每个任务分配给其最早-可用的被需求资源。每个族命名其规范开创性引用。

Solver	优先级依据	规范引用
`earliest-start`	拓扑输入顺序	不适用（恒等基线）
`shortest-processing-time`	最短任务时长优先	Smith (1956)
`longest-processing-time`	最长任务时长优先	Graham (1969)
`earliest-deadline`	最早 deadline 优先（deadline-感知）	Jackson (1955)
`earliest-finish-time`	最早可达完成优先	Topcuoglu et al. (2002)
`minimum-slack`	最小调度松弛优先	Conway, Maxwell & Miller (1967)
`greedy-completion`	最早状态-感知完成优先	Graham (1966)
`apparent-tardiness-cost`	最高表观-延迟-成本指数优先（deadline-感知）	Vepsalainen & Morton (1987)

已退役的派遣族

资源-映射族被显式记录，而非静默地折叠进一个通用类型。它们控制资源分配，而非任务顺序，且尊重-需求的串行构造器在当前核心问题模型下不暴露任何自由映射决策：

资源映射（olb、met、mct、round-robin、load-balanced）：构造器将每个任务分配给其声明的资源，因此这些启发式没有可实现的自由度；执行-时间-感知的映射归属于目标与约束工作，而非此基础。
类型标签缺失（type-aware、domain-aware）：需要当前核心问题模型不携带的任务-类型与资源-类型标签。

retired_dispatching_families() 返回带有每-族理由与引用的完整账本。

基于秩的列表调度器

三个异构-感知的列表调度器共享一个两-阶段形态：一个静态优先级排定阶段依据平均执行时间与通信成本在优先级 DAG 上为每个任务排秩，而一个处理器-选择阶段以优先级顺序通过串行构造器的最早-完成规则绑定每个任务。每-资源执行时间来自一个可塑任务声明的执行模式，而一个刚性任务退化为已记录的同构特化。

Solver	优先级依据	规范引用
`heft`	向上秩	Topcuoglu, Hariri & Wu (2002)
`cpop`	组合的向上-加-向下秩，关键路径任务优先	Topcuoglu, Hariri & Wu (2002)
`peft`	乐观-成本-表秩	Arabnejad & Barbosa (2014)

就绪-集映射规则

三个批量映射规则依据每个优先级-就绪任务在其候选绑定上的最早完成为其评分，然后每步映射一个任务。已发表的规则映射一批独立任务；此处该批即优先级-就绪集，因此这些规则拓展到有依赖的工作负载，并在独立-任务实例上归约为已发表的行为。一旦可塑的 TaskSpec.modes 使其每-机器期望-完成矩阵可表示，这些族便离开已退役账本转为一等注册。

Solver	优先级依据	规范引用
`min-min`	最小最优完成优先	Ibarra & Kim (1977); Braun et al. (2001)
`max-min`	最大最优完成优先	Ibarra & Kim (1977); Braun et al. (2001)
`sufferage`	最大次优-减-最优完成差距优先	Maheswaran et al. (1999)

迭代贪心、禁忌与 RCPSP 基线

三个已发表的单-解方法搜索优先级-安全的任务顺序空间，各自将其规范机制投影到共享的顺序-运行接缝上。serial-sgs-justification 是该平台首个资源-受限项目-调度（RCPSP）求解器。

Solver	机制	规范引用
`iterated-greedy-rs`	NEH 播种，带破坏-重构循环与固定-温度接受	Ruiz & Stutzle (2007)
`critical-path-tabu`	在关键路径块移动上的高级禁忌搜索	Nowicki & Smutnicki (2005)
`serial-sgs-justification`	带双重先右-后左对齐的串行调度-生成-方案解码	Valls, Ballestin & Quintanilla (2005)

元启发式基线

代表性的种子化元启发式基线共享相同的可行性、修复、评分与局部-搜索算子，各自带有其规范引用：遗传算法（Holland 1975）、模拟退火（Kirkpatrick et al. 1983）、蚁群（Dorigo, Maniezzo & Colorni 1996）、粒子群（Kennedy & Eberhart 1995），以及差分进化（Storn & Price 1997）。

编码适配器求解器

综合学习粒子群优化与成功-历史自适应差分进化在一个连续的实-值向量上搜索，每个任务对应一个分量。每个核心对一个连续目标是通用的，因此其收敛行为直接在一个连续基准上验证，并通过一个编码适配器绑定到调度问题，该适配器将一个向量解码为一个优先级-可行的任务顺序。

Solver	灵感	适配	编码
`clpso`	综合学习粒子群（Liang et al. 2006）	连续群解码为优先级顺序	random-key
`d-clpso`	综合学习粒子群（Liang et al. 2006）	在最小-位置-值解码之上的离散适配器	spv
`lshade`	带线性种群缩减的成功-历史自适应 DE（Tanabe & Fukunaga 2014）	连续差分进化解码为优先级顺序	random-key
`d-lshade`	带线性种群缩减的成功-历史自适应 DE（Tanabe & Fukunaga 2014）	在整数-秩取整之上的离散适配器	rounding

CLPSO 从一个综合-学习范例而非单一全局-最优学习每个维度，因此群的最佳适应度在单峰盆地上单调改善。L-SHADE 从其成功历史自适应交叉与尺度-因子记忆，并将种群线性缩减至最少四个个体。两者都报告一条收敛轨迹，且一个收敛-正确性测试在观测轨迹与已发表参考行为相矛盾时失败，独立于种子-可复现性检查。

这些求解器是通过一个编码桥适配到离散域的连续优化器；它们不是任何先前代码的精确重实现，且在比较性活动运行之前不作任何性能声明。

编码适配器

每个具名的连续-到-离散编码都是一个带有显式修复策略的独立适配器，因此没有编码被静默地折叠进一个通用解码器。每个适配器将其解码的顺序修复为一个优先级-可行的顺序，并记录修复是否触发。

适配器	传递	解码规则	引用状态
`random-key`	恒等	按截断键排序	规范（Bean 1994）
`spv`	恒等	最小位置值	规范（Tasgetiren et al. 2007）
`rounding`	恒等	整数秩槽位	不适用
`sigmoid`	S-形 sigmoid	加权概率抽取	规范（Kennedy & Eberhart 1997）
`v-shaped`	V-形幅度	加权概率抽取	规范（Mirjalili & Lewis 2013）
`tanh`	平移双曲正切	加权概率抽取	规范（Mirjalili & Lewis 2013）

确定性解码（random-key、spv、rounding）忽略随机源；传递-函数解码消耗一个种子化源并在固定种子下重放。rounding 适配器记录引用-不适用，因为最近-整数秩取整是一种通用离散化，没有单一规范开创性来源。

多样化竞争者集合

多样化竞争者集合以一小撮近期的强力同侪将比较拓展到代表性基线之外，每个都是一个具名求解器，其全文引用被投影到仅-顺序的决策空间上，而非一份庞杂的经典基线名单。一个同侪仅当其越过一个场所-强度门槛时才留在公共赛场上：发表于后-2020 时效窗口内一个已索引、经同行评审的场所，且场所层级被记录，以便读者基于证据而非声望来权衡它。

Solver	机制	场所	引用
`epso`	带路径汇集的工作负载-偏置-初始化群	Electronics (MDPI), 2023 — 已索引	Anbarkhan & Rakrouki (2023)
`adpso`	带成功-自适应递减惯性的群搜索	Sensors (MDPI), 2022 — 已索引	Nabi et al. (2022)
`ccgp`	优先级-规则树的协同协进化	Computers & Operations Research (Elsevier), 2024 — 顶级 OR	Zaki et al. (2024)

每个同侪在其元数据中携带优势、注意事项，以及一个可运行-基线证据类别，以便推荐器能解释一个求解器为何契合某个上下文。一个更广的经典基线集合——整数-键、偏置-随机-键，以及分布-估计遗传算法，以及大-邻域、迭代贪心、禁忌、可变-邻域，与文化基因搜索——为内部比较保持注册，但被排除在公共赛场之外，因为代表性启发式已携带其机制信号。在比较性活动运行之前不作任何性能声明。

时效锚点

竞争者与相关-工作集合通过一个后-2020 时效锚点针对当前-十年的工作定位：Karimi-Mamaghan、 Mohammadi、Pasdeloup 与 Meyer（2023，通过集成进迭代贪心的 Q-learning 为排列流水车间学习选择算子，European Journal of Operational Research 304(3):1296-1330, doi:10.1016/j.ejor.2022.03.054），以及 2024 IEEE Transactions on Evolutionary Computation 为分布式流水车间组调度提出的指标-驱动协作多-目标进化算法（doi:10.1109/TEVC.2023.3339558）。

多-目标竞争者

两个 Pareto 竞争者在一个多-目标向量上演化一个优先级-安全任务顺序的种群，区别于众-目标 nsga3 表面。nsga2 是基于支配的竞争者——快速非支配排序，带拥挤-距离平局决胜——而 moead 是基于分解的对衡，沿一个结构化单纯形权重网格将问题拆分为标量 Tchebycheff 子问题，并在每个子问题的最近-权重邻域上替换在位者。

Solver	机制	规范引用
`nsga2`	基于支配的 Pareto：快速非支配排序、拥挤距离	Deb et al. (2002)
`moead`	基于分解的 Pareto：单纯形权重网格上的 Tchebycheff 标量化	Zhang & Li (2007)

众-目标竞争者

nsga3 是具名的众-目标竞争者（Deb & Jain 2014, doi:10.1109/TEVC.2013.2281535），区别于双-目标竞争者表面。它在优先级-安全的任务顺序上维护一个种群，在一个多-目标向量（makespan、延迟与负载公平性）上评估每个顺序，并通过在 Das & Dennis 单位单纯形上的结构化参考点之上的参考点-小生境选择，使每一代存活。参考点设计、解到其最近参考方向的关联，以及小生境-计数选择是算法签名；该求解器在 multi-objective 之外声明 many-objective 能力，以便一个活动可显式地选择它。

能力匹配

SolverRegistry.select 按能力、约束与目标匹配求解器。能力与约束要求都表达为能力标签并作为合取匹配：一个同时需要 deadline-感知的众-目标活动传入 required_capabilities=(SolverCapability.MANY_OBJECTIVE,) 与 required_constraints=(SolverCapability.DEADLINE_AWARE,)，注册表仅返回同时声明两者的求解器。一个 objective 过滤器进一步将结果限制为声明支持该具名目标的求解器。

NDSO 族

NDSO 族是注册表的原生-编码求解器族：它直接在可行调度上搜索。它产生的每个调度都按构造可行（一个有效性-按-设计的构造器在每一步构建一个尊重-优先级的顺序），因此该族携带 native 编码，且从不运行编码/解码步骤或修复过程。该族组合一小组具名机制：

信心矩阵 — 一个按（位置、任务）单元存储已学习信任的稀疏存储，随着更好的调度强化其单元而更新。
信心-加权投票 — 通过在由信心矩阵加权的种群上投票来合成精英调度；快速变体改用一个非-加权多数投票。
数量-与-质量调度 — 数量调度设定一个候选改变多少；质量调度设定它从哪个指导源学习。
三-源指导 — 一个候选从合成精英（开发利用）、一个同侪（多样性），或一个消逝-知识源（激进探索）学习。
统一自适应系数 — 一个非-线性调度将该族从探索转向开发利用，并驱动敏感性与学习焦点二者；快速变体将其固定为一个固定值。

变体	组成
`ndso-core`	信心-加权投票、自适应系数、三-源指导
`ndso-fast`	多数-投票合成、固定系数、单一指导源
`ndso-summit`	协调若干群并带总体合成的群间-议会

群间-议会

ndso-summit 变体是质量组成：它并行运行若干群并通过一个议会协调它们。每个群在其自己的种群上组合核心机制；群间-议会使那些群作为一次搜索而非若干孤立运行来工作，并将其结果合成为一个单一的总体精英——整个峰会所支持的调度。议会是一眼区分该变体之处：ndso-core 与 ndso-fast 各以一个种群搜索，而 ndso-summit 是为协调的多-群搜索构建的组成。

议会的协调行为与每-群诊断是可配置的，且它产生的每个调度都按构造保持可行。

消融图

消融图为每个具名机制枚举一个隔离配置，以便下游分析能归因每个机制的贡献。群内条目各自禁用一个核心开关——信心矩阵、信心-加权投票、数量调度、质量调度、多-源指导结构、同侪源、消逝-知识源，以及自适应系数。协调条目各自禁用一个议会参数——群内- 对群间-协调，以及总体合成。每个条目声明分析层采样其比较所在的运行-计数下限，并命名该层应用的统计检验（一个配对非-参数显著性检验、一个带 Holm 多重-比较校正的 Friedman 平均-秩事后检验，以及一个 Cliff-delta 效应量）。被采样的比较由下游实验引擎物化；进程内 runner 证明每个隔离都可行。

一个分阶段导出过滤器治理一个报告范围可暴露哪些机制：基础范围仅暴露基础性机制，且一个 fail-closed 门在基础范围下抛出而非泄露一个更-快或更高-质量的机制，因此基础与快速范围无法暴露仅-议会机制。收敛诊断——种群多样性、信心-矩阵熵、探索比率、资源负载、计时，以及改善轨迹——是可选的且在禁用时不增加开销，且议会导出一个分析层无需导入任何求解器类型即可摄入的 JSON 清单。

精确适配器

精确适配器采取两种形态之一。可选-后端适配器声明一个可选依赖并在运行时检查其导入根；若后端缺失，它们抛出 MissingOptionalDependencyError，带有 extra 名称、用途、商业标志与许可注记，而非在包导入期间导入一个重量级依赖。原生-有界适配器不携带第三方依赖，并通过其自身的有界搜索 ——一者枚举优先级-可行顺序，另一者为决策图分支定界——精确求解小实例，在实例超过所支持的任务计数时抛出 UnsupportedCapabilityError。

Solver	方法	后端	Extra	商业
`ortools-cp-sat`	CP-SAT	`ortools`	`exact`	否
`pulp-milp`	MILP / MIP	`pulp`	`exact`	否
`branch-and-bound`	branch-and-bound / cut	`ortools`	`exact`	否
`logic-based-benders-decomposition`	基于逻辑的 Benders 分解	`pulp`	`exact`	否
`exhaustive-enumeration`	穷举枚举	原生（无后端）	—	否
`decision-diagram-sequencing`	决策图分支定界	原生（无后端）	—	否
`gurobi-exact`	MILP / MIP	`gurobipy`	`exact-commercial`	是

公共的精确一小撮是代表性开放集合——CP-SAT、MILP、branch-and-bound 与基于逻辑的 Benders 分解——加上两个原生有界求解器与真正调用的 Gurobi 封装。商业 Gurobi 封装留在 exact-commercial extra 之后，从不被捆绑，并携带一个显式许可注记；有一个学术许可可用。一个更广的可选后端池为内部使用保持注册。

选择与学习层

选择层为一个问题对求解器排名，而从不声称一个全局-最佳求解器。它消耗一个 SelectionFeatures 记录——difficulty、heterogeneity、objective_conflict、uncertainty、dynamism，以及 solver_sensitivity——与公共求解器元数据，并返回排名的 SolverRecommendation 行。每个推荐携带一个 RecommendationSource（metadata 或 learned-model）、一个 ConfidenceLabel，以及显式限制注记，因此一个元数据推荐绝不会被误认为一个已学习的推荐。

FEATURE_ORIGIN 将每个特征追溯到它读取的基准特征化度量；消费者将 dispatchatlas.bench 的特征化度量映射进特征契约，因此 dispatchatlas.solve 仍仅导入 dispatchatlas.core。两个选择器发布：RuleBasedSelector 仅从已声明能力与特征化排名（一个元数据推荐），而 SupervisedSelector 从一个带标签语料库通过一个确定性距离-加权最近-邻模型排名（一个已学习-模型推荐）。

有监督选择器报告留出泛化，而非训练拟合。交叉-验证协议划分一个带标签语料库，使得没有实例、没有基准族，且没有特征化记录同时出现在训练与测试划分中（partition_by_families、 leave_one_family_out、leakage_report），在剩余族上训练，并对留出族评分（held_out_generalization、cross_validate）。其信心仅当一个无-泄露的留出评估支持它时才升至元数据-唯一之上。

learning_interface_catalog() 注册七个具名的学习与混合接口——有监督算法选择、代理-辅助搜索、强化-学习钩子、超-启发式、策略-引导修复、已学习初始化，以及一个仅-基准基线。每个声明一个 LearningEvidencePolicy（训练数据、泄露控制、可复现性、证据类别、证据-层级资格）。两个在此实现；其他五个是已注册的延迟接口，其实现取决于产生训练数据的比较性活动。重型估计器后端留在可选的 learning extra 之后并按导入根探测，从不在包导入期间导入；缺失抛出 MissingOptionalDependencyError，而确定性回退保持可用。

性能保障

批量评分通过 BatchScoringProfile 显式化。当前内核使用一个标准-库回退并在诊断中记录该回退。这使 API 为向量化或加速内核保持就绪，同时保留一条经测试的、可移植的路径。