Tracking the Race Between Deep Reinforcement Learning and Imitation Learning -- Extended Version

TL;DR

This paper compares deep reinforcement learning and imitation learning on a Racetrack benchmark, showing reinforcement learning produces more foresighted agents while imitation learning leads to riskier paths, highlighting their differing decision-making characteristics.

Contribution

The study provides a comparative analysis of deep reinforcement learning and imitation learning on a benchmark problem, revealing their distinct behaviors and performance traits.

Findings

Reinforcement learning agents are more foresighted and avoid risky states.

Imitation learning agents tend to follow more risky paths.

Reinforcement learning generally outperforms imitation learning in this task.

Abstract

Learning-based approaches for solving large sequential decision making problems have become popular in recent years. The resulting agents perform differently and their characteristics depend on those of the underlying learning approach. Here, we consider a benchmark planning problem from the reinforcement learning domain, the Racetrack, to investigate the properties of agents derived from different deep (reinforcement) learning approaches. We compare the performance of deep supervised learning, in particular imitation learning, to reinforcement learning for the Racetrack model. We find that imitation learning yields agents that follow more risky paths. In contrast, the decisions of deep reinforcement learning are more foresighted, i.e., avoid states in which fatal decisions are more likely. Our evaluations show that for this sequential decision making problem, deep reinforcement learning performs best in many aspects even though for imitation learning optimal decisions are considered.