Hyperbolic Discounting and Learning over Multiple Horizons
This addresses a fundamental modeling issue in RL for aligning with biological and economic insights, with incremental improvements in agent performance.
The paper tackles the disconnect between reinforcement learning's exponential discounting and evidence of hyperbolic time-preferences in humans and animals by implementing an RL agent with hyperbolic discounting, showing that a simple approximation works with standard techniques, and discovers that learning value functions over multiple horizons improves performance over the Rainbow agent.
Reinforcement learning (RL) typically defines a discount factor as part of the Markov Decision Process. The discount factor values future rewards by an exponential scheme that leads to theoretical convergence guarantees of the Bellman equation. However, evidence from psychology, economics and neuroscience suggests that humans and animals instead have hyperbolic time-preferences. In this work we revisit the fundamentals of discounting in RL and bridge this disconnect by implementing an RL agent that acts via hyperbolic discounting. We demonstrate that a simple approach approximates hyperbolic discount functions while still using familiar temporal-difference learning techniques in RL. Additionally, and independent of hyperbolic discounting, we make a surprising discovery that simultaneously learning value functions over multiple time-horizons is an effective auxiliary task which often improves over a strong value-based RL agent, Rainbow.