Deep MARL Agents Collude in Asynchronous Pricing Markets Despite — and Because of — Asynchrony
Deep multi-agent RL (DDPG) agents tasked with competitive pricing in a continuous-time market reliably form tacit cartels (collusion index 0.69±0.11), raising the price floor above the Bertrand equilibrium. Introducing asynchrony in price update times reduces collusion by 48%, but the effect is non-monotone in observation latency and breaks down entirely under critic divergence at high event rates (λ=5). The paper provides a reproducible benchmark and trajectory-level diagnostics for detecting within-episode signalling collapse and post-shock non-recovery, but the proposed 'fix' is partial and fragile.