Kaushik Roy: “Relational Reinforcement Learning”

less than 1 minute read

My work is mostly about approaches to solving decision making problems in Relational domains. The main shortcoming about these methods is that to do function approximation you would need to sample far too many samples to be able to perform a decent global approximation using a machine learning method. However, because of the potentially infinite state space, we employ gradient boosted methods that induce interactions as necessary and solve this problem efficiently.

The work has been applied to both propositional and relational domains and the results have been accepted at AAMAS 2018 under the topic of “Non-parametric Fitted Relational Value Iteration: Unifying Relational and Propositional Discrete Domains”. I continue to work towards efficient RL methods in relational domains, the goal is to solve a problem in the POMDP (partially-observable Markov decision process) setting.