Transformer based Meta-Reinforcement Learning for Dynamic Cancer Treatment

The field of Meta-Reinforcement Learning has gained a lot of traction in recent years, especially in the overarching research concerning control and decision making. However, real life applications of these methodologies are far and few in between.. In our work we seek to expand as well as have a fresh take at the dynamic treatment regime problems and the existing deep meta-RL solutions to it. We attempt to do so by incorporating the very promising Transformer-encoder architecture that has practically revolutionized speech and image processing. We propose a new architecture that effectively exploits the self-attention mechanism present as a contextualization framework that acts as a memory component for the meta-algorithm. We host a policy gradient methodology above this meta-framework that solves sequential decision problem for treating Head and Neck Cancer patients.