About

Martin Balla is currently a PhD student in the Game AI group at the Queen Mary University of London as part of the Intelligent Games and Game Intelligence Doctoral Training (IGGI) programme. Martin’s thesis focuses on multi-task learning using Successor Features (SFs) and Goal-conditioned RL with the objective of transferring across various reward functions with less interactions. During his PhD Martin looked into General Video Game Playing using Deep RL across multiple games/levels and visuals, and Goal-conditioned RL agents using Universal Value Function Approximators (UVFAs) with Hindsight Experience Replay (HER) to make them more efficient. The main contributions of his work are around Successor Features combined with various extensions, including task relabelling and methods to learn various policies for transferring.

Martin is interested in Reinforcement Learning agents that can adapt to changes in the reward function and/or changes in the environment. His current work focuses on learning Hierarchical RL policies end-to-end using SFs. Outside of his main research direction, Martin is involved with the Tabletop games framework (TAG), which is a collection of various tabletop games implemented with a common API with a focus on various game-playing agents (including RL). TAG brings various challenges to RL agents compared to search-based agents, such as complex action spaces, unique observation spaces (various embeddings), multi-agent dynamics with competitive and collaborative aspects, and lots of hidden information and stochasticity.