Risk-sensitive Distributional Reinforcement Learning
Reinforcement learning (RL) is a powerful framework for training agents to maximize cumulative reward, but it typically assumes risk-neutrality. This can lead to suboptimal behavior in practical scenarios where the consequences of unfavorable outcomes can be detrimental. What is risk? Generally, risk might arise whenever there is uncertainty. In a financial situation, investment risk can be identified with uncertain monetary loss. In a safety-critical engineering system, risk is the undesirable detrimental outcome....