Abstract
A computer-implemented audit evaluates collusion risk in multi-agent reinforcement-learning pricing systems by analyzing the full counterfactual structure of trained policies rather than only observed price traces. Trained policies are frozen into deterministic greedy action mappings over a discrete joint market state space. A deterministic successor mapping is computed for each joint state to form a directed functional graph with one outgoing edge per state. The graph is decomposed into attractor cycles and basins of attraction that partition the state space. Each attractor is assigned an elevation score based on mean cycle price relative to reference Nash and joint-profit-maximizing price levels, classified by configurable thresholds, and weighted by basin mass. A basin-weighted Attractor Collusion Risk score aggregates excess elevation above a threshold. Optional analysis includes Q-value forensics via static best response comparison and state-information ablation comparing full-state and no-state training conditions.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Policy-Graph Audit Method for Algorithmic Collusion Detection in Multi-Agent Pricing Systems", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10944