Shreya Sinha Roy

I am a 4th year PhD student working under the supervision of Dr. Ritabrata Dutta, Dr. Richard Everitt, and Prof Christian Robert. My primary research interests include computational statistics, Bayesian inference, sampling for high-dimensional parameter spaces, and reinforcement learning.

Bayesian Deep Generative Reinforcement Learning:

Bayesian deep RL: The above diagram shows an episodic posterior update followed by a policy update routine for $j$ th episode $j=1,2, ...$ . We have assumed the episode length to be $tau$ , and $pit$ denotes the prior on the model parameters. Prequential Scoring Rule ( $s$ ) is used to compute the generalized posterior which is based on true interaction data( $x$ ) and a simulation( $xt$ ) from the deep generative model $m$ . We obtain the samples, $thetas$ from the generalized posterior, $post$ via Sequential Monte Carlo (SMC) samplers. These samples are used to simulate $n$ trajectories of interaction $sim$ from the model $m$ . Optimal policy is trained by maximizing the averaged value of $q$ function computed from the $n$ simulated trajectories. The new policy, $mu$ is then used to interact with the true Environment in the next episode.