Answers are grounded in this paper and cite their sources.
Reinforcement learning (RL) remains a central optimization framework in machine learning. Although RL agents can converge to optimal solutions, the definition of ‘optimality’ depends on the environment's statistical properties. The Bellman equation, central to most RL algorithms, is formulated in terms of expected values of future rewards. However, when ergodicity is broken, long-term outcomes depend on the specific trajectory rather than on the ensemble average. In such settings, the ensemble average diverges from the time-average growth experienced by individual agents, with expected-value formulations yielding systematically suboptimal policies. Prior studies demonstrated that traditional RL architectures fail to recover the true optimum in non-ergodic environments. We extend this analysis to deep RL implementations and show that these, too, produce suboptimal policies under non-ergodic dynamics. We show that this limitation can be addressed by treating non-ergodic decision problems as finite trajectory-exposure problems rather than as isolated one-step optimization tasks. By exposing the agent to compounded trajectories during training, the learned value estimates become more consistent with the process's intrinsic time growth. This improvement does not require altering the environmental feedback, such as reward transformations or modified objective functions, but arises naturally from the agent's exposure to temporal trajectories. Our results contribute to the growing body of research on RL methods for non-ergodic systems. This article is part of the theme issue ‘Ergodicity and ergodicity breaking in the sciences’.
No takes yet. Share an insight, caveat, or question.
Verbruggen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: