{"id":"fc56f5a2-13ce-4fd8-99dd-177de90a7c57","arxiv_id":"2506.20404","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A software library, GymPN, extends the A-E Petri net framework with partial observability and multiple action transitions, and learns optimal task assignment policies on eight workflow patterns.","lead":"GymPN is a new Python library that helps businesses decide, in real time, which worker should do which task by combining process models with deep reinforcement learning. It extends an earlier framework to handle partially visible process information and multiple decision points, and the authors show it learns optimal assignment policies on eight benchmark patterns.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Expansion step lacks a dynamics-preservation argument, and the eight benchmarks never exercise a multi-token place, so the general optimal-policy claim is not yet supported.","rationale":"The reader's weakest assumption identified the expansion-to-single-token issue, and my reading of the full text confirms it is load-bearing. Algorithm 1 is the only mechanism translating a Petri net state into an observation, and it is undefined unless each place holds at most one token. The paper never provides the expansion definition or a preservation proof; the example figures are illustrative only. Because every evaluation problem keeps at most one token in each place at every decision point, the experiments cannot distinguish a correct mapping from one that fails on a simple backlog scenario. This is a concrete, addressable gap rather than a fatal flaw, so conditional acceptance is the right verdict: the authors should supply the missing preservation argument or explicitly restrict the library's scope, and add a multi-token experiment. I do not see a separate objection that would warrant rejection; the claim about the eight tested patterns is plausible as far as it goes, and the unevaluated partial-observability feature is disclosed as a current limitation rather than a contradiction of the main evaluation.","tokens_in":11451,"tokens_out":2334,"duration_ms":23093,"concrete_test":"Build a minimal two-case benchmark: an A-E PN with a single Waiting place initially holding two case tokens of different task types and one resource place holding two resource tokens, arranged so both cases are queued when the action phase starts. Run the documented expansion procedure and Algorithm 1, then compare the original net and the expanded net on every deterministic policy: check that the set of enabled bindings is identical and that firing each binding yields the same successor marking, clock, and reward. Also inspect the assignment graph to verify it contains two distinct nodes for the two waiting cases. If the expansion changes the enabled set or the graph collapses the two tokens into one node, the mapping is not semantics-preserving and the current evaluation does not cover the library's claimed general scope.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Algorithm 1 assumes every place holds at most one token: the mapping in Section 4.2 says 'every non-empty place is mapped to a node with the same attributes as the single token in the original place', and the loop uses p(t) for 'a place p marked with a single token t'. Plain A-E PN places hold multisets of tokens; fig. 1 already has multiple cases in Waiting. The paper only says 'an expansion procedure' produces the one-token-per-place form and shows two example expansions (figs. 6 and 9), but never defines this procedure or proves it preserves enabled transitions, bindings, rewards, and timing. Without that preservation argument, the observation sent to the agent may misrepresent the true process state, exactly the concern flagged by the reader. The eight evaluation patterns do not stress this gap: each problem has at most one active task per decision point, so every place carries at most one token whenever an action is chosen; arrivals are immediately assigned and resources are three static tokens. Thus the experimental evidence only exercises the trivial single-token case of Algorithm 1 and cannot detect a failure of the expansion on a genuine multi-token marking. Additionally, the partial-observability novelty is never evaluated, though the headline claim rests mainly on the multiple-action results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GymPN, a Python library that extends the Action-Evolution Petri Net (A-E PN) framework with two features: partial observability of process state, realized through place- and attribute-level observability masks, and support for multiple action transitions, realized via a one-hot encoding of transition types and a mapping from an expanded Petri net to an assignment graph. The library is evaluated on eight task-assignment instances covering sequence, parallelism, arbitrary cycle, and exclusive-choice workflow patterns with joint or disjoint resource pools; the paper reports that a PPO agent reaches the hand-computed optimum reward in all eight instances. The library is released as open source, and the paper includes a comparison with related decision-making frameworks.","tokens_in":11713,"tokens_out":5005,"duration_ms":56468,"significance":"If the claims hold, GymPN addresses a real gap: previous A-E PN implementations assume full observability and a single decision type, while real business processes often require decisions at multiple stages and may hide part of the state. The open-source release, the clearly specified eight benchmark patterns, and the exact match of the reported PPO rewards to the stated optima in Table 2 are strengths that give some confidence in the multi-action encoding for the tested cases. However, the significance is currently conditional: the partial-observability novelty is only defined and never evaluated, and the multi-action mapping relies on an unformalized and unproven expansion step that the experiments do not stress. These gaps need to be closed before the paper's central claims are fully supported.","major_comments":[{"comment":"Algorithm 1 and Definition 4 are stated for an 'expanded' A-E PN in which every place holds at most one token, but the expansion procedure itself is never formally defined and no proof is given that it preserves the set of enabled transitions, their bindings, firing times, or rewards. The paper only shows two example expansions (Figs. 6 and 9) and a short informal paragraph. Without a preservation argument, the assignment graph may misrepresent the true process state, and a policy learned on the expanded net may be suboptimal or invalid for the original multi-token net. This is load-bearing because the multiple-action claim rests on the faithfulness of this mapping.","section":"§4.2, Definition 4, Algorithm 1"},{"comment":"The eight benchmark patterns do not, as described, exercise a genuine multi-token decision point: under the optimal deterministic policy, arrivals are assigned before queues build, and the resource pools are static token sets. Thus the empirical results validate Algorithm 1 only in the trivial single-token-per-place case and cannot detect a failure of the unproven expansion on a multi-token marking. The authors should either provide a formal proof of expansion correctness or add a benchmark with a congested queue where multiple waiting tokens are simultaneously visible at a decision point, and verify the learned policy against an independently derived optimum.","section":"§6, Table 2"},{"comment":"Partial observability, one of the two headline novelties, is never evaluated. Section 4.1 defines the observability functions O_P and O_A, but Section 6 evaluates only fully observable task-assignment problems. There is no experiment that masks a place or attribute, no comparison between a partially observable agent and a fully observable baseline, and no discussion of how the action space or the optimal policy changes under hidden information. The claim that GymPN 'supports partial process observability' is therefore only a modeling claim, not a demonstrated capability.","section":"§4.1, §6"},{"comment":"No training hyperparameters or network architecture are reported. The paper states only that PPO is used and that graph neural networks approximate the policy and value functions; it gives no learning rate, batch size, number of training steps, GNN architecture, or number of random seeds. With only 10 evaluation episodes per problem and no variance across independent training runs, the exact equality of PPO and optimum rewards in Table 2 is not sufficient to establish that GymPN reliably learns optimal policies. A reproducibility appendix with these details would substantially strengthen the paper.","section":"§5, Table 2"}],"minor_comments":[{"comment":"In the joint-action example, the text says the expansion produces Fig. 6 and the mapping produces Fig. 7, but the joint expansion and its assignment graph are shown in Figs. 9 and 10; both references should be corrected.","section":"§4.2, Figs. 6–10"},{"comment":"The tuple in Definition 1 has a trailing comma and a stray 'A' after the closing parenthesis ('..., ρ0), A'); this formatting error should be fixed.","section":"Definition 1"},{"comment":"The package is referred to as 'POMDPs.jl' in the running text and 'POMDP.jl' in Table 1; the name should be used consistently.","section":"§2, Table 1"},{"comment":"The outline in Section 1 lists 'section 6' twice and repeats the evaluation description; the section numbering and outline should be tightened.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"This is a software paper with a clear niche, and the gaps identified above are fixable within the manuscript's scope. I would support acceptance if the authors add a rigorous preservation argument for the expansion step, evaluate partial observability with at least one concrete partially observable problem, and report training details. If the expansion cannot be proved or tested, the multiple-action claim should be explicitly restricted to one-token-per-place nets."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on 2506.20404.\n\nThis is a real systems contribution, but it's narrower than the abstract makes it sound. The library extends the authors' A-E PN framework with two features: partial observability and multiple action transitions. The multiple-action part is tested on eight workflow patterns and works: the PPO policies match the hand-derived optima exactly. That's genuine evidence that the mapping from A-E PN to assignment graph, and the one-hot encoding of action types, is sound for those toy problems.\n\nThe partial-observability feature, though, is only defined, never evaluated. There's not a single experiment where you hide a place or attribute and check that the agent still learns something sensible. So the second headline claim is unsupported.\n\nThe bigger concern is Algorithm 1. It maps an 'expanded' A-E PN in which every place holds at most one token. The expansion that produces this form is described in one paragraph and never formally defined, and there's no argument that it preserves the enabled transitions, bindings, or rewards of the original multi-token net. The eight benchmarks never have more than one token in any place at the moment an action is chosen – arrivals are immediately assigned and resources sit as static tokens – so they can't detect a failure of that expansion. For a framework whose definition allows multisets, this is a real gap. I don't think it's fatal; it's an addressable omission. But as written, the general claim that GymPN can model realistic processes isn't supported.\n\nYou also get no training hyperparameters or network architecture, which is a reproducibility problem for a library paper. And the text has copy-paste errors: figures 6 and 9 get mixed up, and the intro lists section 6 twice. Minor, but sloppy.\n\nWhat the paper does well: it lowers the barrier for process-management researchers to use DRL. You define the A-E PN in terms of places, transitions, guards, rewards; the library builds the Gym environment and handles observations and actions. That's genuinely useful, and the evaluation, though narrow, is honest: they report the known optima and the learned policies match.\n\nBottom line: this deserves serious review. Send it, but expect a major revision. The authors need to (1) run at least one partial-observability experiment, (2) either prove the expansion preserves dynamics or define it formally and test it on a multi-token case, and (3) report training details. Then it would be a solid contribution for the BPM community.","headline":"Useful library paper with one well-tested feature and one untested feature; the multi-token expansion needs a preservation argument.","tokens_in":12194,"tokens_out":3501,"would_cite":false,"duration_ms":35787,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GymPN models business-process decisions as partially observable Action-Evolution Petri nets, and its DRL agent learns the optimal assignment policy on all eight benchmark patterns.","keywords":["Process Models","Decision Making","Reinforcement Learning","Process Management Systems","Petri Nets","Graph Neural Networks","Partial Observability","Action-Evolution Petri Nets"],"falsifier":"Take a small A-E Petri net in which a single place holds two tokens and one action transition consumes both tokens at once (or has a guard depending on both), expand the net as Algorithm 1 does, and check whether the expanded net enables the same transitions with the same timing and rewards. If the enabled bindings, clock updates, or reward sequence differ from the original net, the assignment graph misrepresents the true process state and a policy trained on it cannot be trusted as optimal.","tokens_in":11260,"feed_emoji":"🎯","tokens_out":9197,"duration_ms":82071,"temperature":0.7,"pith_summary":"GymPN is a software library that turns business-process decision problems—which task to do next, when, and who should do it—into a Petri-net model that a deep reinforcement learning agent can solve. The paper's central claim is that the library's two new features, partial observability and multiple action transitions, let a single framework represent realistic process decisions that earlier Action-Evolution Petri net implementations could not. On eight benchmark problems covering the standard workflow patterns (sequence, parallelism, arbitrary cycle, and exclusive choice, each in joint and disjoint variants), a PPO agent learned policies that match the known optimum in every case. If correct, this gives process managers a practical, open tool for finding optimal allocation policies rather than relying on hand-tuned heuristics.","feed_headline":"New RL library hits optimal assignment on all eight test patterns","feed_subtitle":"Adds partial observability and multi-decision modeling; PPO matches the optimum in all eight benchmarks.","key_machinery":"The central object is the assignment graph, a graph-structured observation that the simulator extracts from the current marking of an Action-Evolution Petri net (A-E PN)—a timed, colored Petri net whose 'action' transitions are the decisions an agent makes and whose 'evolution' transitions are the nondeterministic events of the process. The mapping of Definition 4 and Algorithm 1 builds this graph only from observable places and attributes: each non-empty observable place becomes a node carrying the token's observable attributes, each transition becomes a node with a one-hot encoded type so multiple action transitions are distinguishable, and arcs between observable elements become edges. The two new observability functions, $O_P$ and $O_A$, decide which places and which token attributes are visible, and the unobservable elements are simply omitted. The graph is fed to a graph neural network that selects an action node, so the agent's policy operates on a partial view of the true process state.","core_discovery":"The paper presents GymPN, an extension of the Action-Evolution Petri net (A-E PN) framework for business-process decision making, and claims two novelties: it introduces observability functions that can hide entire places or individual token attributes from the agent's assignment-graph observations, and it extends the mapping from Petri net to assignment graph so that multiple action transitions—whether they share resources (joint actions) or use separate pools (disjoint actions)—are represented and distinguished by one-hot encoded transition types. To do this, the net is first expanded so that every place holds at most one token, then mapped to a graph whose nodes and edges carry only the observable places, attributes, and transitions. The paper evaluates the library on eight problem patterns and reports that a trained PPO agent achieves the optimal cumulative reward in all of them, with zero variance on the deterministic problems.","pith_inferences":["The paper's evaluation hides the arrival process from the agent in the partial-observability example, but all eight benchmarks are effectively fully observable; a stronger test of the novelty would hide a token attribute that changes the optimal choice and check whether the agent learns a policy that compensates for the missing information.","The correctness of the learned policies ultimately rests on the unproved assumption that expanding multi-token places into single-token places preserves the original net's enabled transitions, bindings, and timing; a directed test of that preservation on a net with a transition that consumes two tokens from one place would either validate or break the library's core mapping.","The one-hot encoding of action transitions means the observation graph grows with the number of decision points; scaling to processes with dozens of action transitions may require a compact action representation, a question the paper does not address."],"forward_implications":["Process analysts can model partially observable decision problems—for instance hiding the arrival process from the agent—without switching to a different formalism, because observability is expressed as two functions on places and attributes.","A single trained agent can handle processes with several independent decision points, including cases where two actions draw on the same resource pool, because action transitions are one-hot encoded in the observation.","The same library API covers all four basic workflow patterns, so the reported result suggests that optimal assignment policies can be learned across a wide range of process structures rather than for one bespoke formulation.","Because the deterministic policy attains the computed optimum in every benchmark, the method is a candidate replacement for hand-coded heuristics in settings where the optimal policy is known to exist in the modeled class."],"supporting_citations":[{"why":"SimPN, the Python Petri-net simulation library that GymPN builds on and extends.","marker":"[4]"},{"why":"Defines the original Action-Evolution Petri net framework and its alternating action/evolution semantics, which GymPN extends.","marker":"[10]"},{"why":"Introduces the assignment graph and graph-neural-network feature representation that Algorithm 1 modifies and uses to expose multiple action types.","marker":"[11]"},{"why":"Gymnasium, the standard reinforcement-learning environment interface that GymPN's gym environment wraps.","marker":"[17]"},{"why":"Workflow patterns, the taxonomy of sequence, parallelism, arbitrary cycle, and exclusive choice that defines the eight evaluation problems.","marker":"[18]"}],"fun_headline_variants":["RL library for process decisions hits optimum on 8 patterns","GymPN: RL for processes, partial observability, multi-action optimal","Process decision RL library: optimal on eight benchmarks","GymPN enables multi-decision RL for processes with partial observability","PPO attains optimal rewards on all eight process-decision patterns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that expanding a net in which a place may hold several tokens into a net in which each place holds at most one token preserves exactly which actions are possible, with the same timings and bindings; the paper assumes this preservation rather than stating or proving it.","fun_headline_variants_meta":{"raw":{"variants":["RL library for process decisions hits optimum on 8 patterns","GymPN: RL for processes, partial observability, multi-action optimal","Process decision RL library: optimal on eight benchmarks","GymPN enables multi-decision RL for processes with partial observability","PPO attains optimal rewards on all eight process-decision patterns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3648,"prompt_tokens":862,"completion_tokens":2786,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":2698}},"tokens_in":478,"tokens_out":2786,"duration_ms":19632,"temperature":1.0,"reasoning_tokens":2698,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:48:32.686799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small A-E Petri net in which a single place holds two tokens and one action transition consumes both tokens at once (or has a guard depending on both), expand the net as Algorithm 1 does, and check whether the expanded net enables the same transitions with the same timing and rewards. If the enabled bindings, clock updates, or reward sequence differ from the original net, the assignment graph misrepresents the true process state and a policy trained on it cannot be trusted as optimal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SimPN, the Python Petri-net simulation library that GymPN builds on and extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the original Action-Evolution Petri net framework and its alternating action/evolution semantics, which GymPN extends."},{"cited_title":"In: Marrella, A., Resinas, M., Jans, M., Rosemann, M","cited_arxiv_id":null,"evidence_quote":"Introduces the assignment graph and graph-neural-network feature representation that Algorithm 1 modifies and uses to expose multiple action types."}],"review_version":1}