Risk-Sensitive Reinforcement Learning

Course Project: Reinforcement Learning (Monisha Jegadeesan, Shreyas Chaudhari)

The following risk-measures and algorithms are implemented for a gridworld with error states setting:

Risk-Neutral

• Q-Learning
• SARSA

Risk as the Probability of Entering an Error State

• R-Learning

Risk as the Euclidean Distance from the Nearest Error State

• S-Learning

The mapping between the risk measures, algorithms and code files can be found in algorithm_file_mapping.txt.

A detailed report on the analysis of these measures and algorithms can be found here.