Development and Analysis of Reward-Adaptive Reinforcement Learning Agents
This research is attempting to develop an artificially intelligent reinforcement learning agent that can adapt to changes in its reward structure without performing training again. An agent has been developed that successfully plays Atari 2600 games with random reward elements. More powerful reward-adaptive agents (RWA) may inform medical scientists of optimal steps to counteract the emergence of drug resistances in several diseases, and other diminishing returns in patient treatment. The next step involves training and testing the RWA in more complex 3-dimensional environments with larger action spaces and observation spaces.