Trace-Mediated Peak Bias: Bridging Temporal Credit Assignment and Cognitive Heuristics in Deep Reinforcement Learning

2026-06-03

AI Summary

This work investigates the interaction between temporal credit assignment and nonlinear function approximation in deep reinforcement learning, which can induce systematic biases that lead agents to favor trajectories with high reward peaks over those yielding greater cumulative returns. For the first time, the paper connects the psychological peak-end rule from human cognition to the optimization dynamics of deep RL, revealing a peak-preference phenomenon caused by gradient surges at deeper eligibility traces. Through a synthesis of eligibility traces, temporal-difference learning, and adaptive optimization techniques—particularly second-moment normalization—theoretical analysis and experiments demonstrate that fixed-stepsize optimizers systematically overestimate global value estimates, whereas adaptive optimizers effectively mitigate this bias, yielding value predictions that better align with rational expectations.

Abstract

Temporal credit assignment is central to both biological and artificial intelligence, yet its interaction with non-linear function approximation is poorly understood. We identify a systematic failure mode in deep reinforcement learning (RL) termed Trace-Mediated Peak Bias (TMPB). At intermediate eligibility trace depths, agents irrationally prefer trajectories with high-magnitude reward ``peaks'' over alternatives with higher cumulative returns. This provides a mechanistic account of the Peak-End Rule: a human memory bias where experiences are judged by their most intense moments rather than integrated utility. We show that TMPB emerges because traces amplify distal Temporal Difference errors into ``gradient shocks'' that fixed-step-size Stochastic Gradient Descent cannot normalize, leading to global overestimation. Conversely, adaptive optimizers mitigate this pathology via second-moment normalization. Our results suggest that human-like saliency distortions may emerge naturally from the mathematical constraints of credit assignment in distributed systems, and that adaptive optimization is a theoretical necessity for rational value estimation.

Problem

• Temporal credit assignment
• Trace-Mediated Peak Bias
• Deep Reinforcement Learning
• Peak-End Rule
• Non-linear function approximation

Innovation

• Trace-Mediated Peak Bias
• Temporal Credit Assignment
• Eligibility Traces
• Adaptive Optimization
• Peak-End Rule