Ádám Katona

Ádám Katona

Intrinsic Motivation in Reinforcement Learning

The aim of this research is to look for ways to automatically discover reward functions, that use the internal state of the learning agent. In nature there is no experimenter to define an external reward function, all learning must be driven by internal signals. In computer science, we do not have such a restriction, however using these internal signals might have some advantages.

One such advantage is the ability to use the abstractions already learned by the agent to express the reward function. This can potentially make it easier for an evolutionary search to discover useful reward functions, to bias learning toward more successful or interesting trajectories.

This kind of reward function discovery might be useful to deal with sparse reward environments. Another application is to use the search to create more interesting or believable agents instead of just competent ones. A way that this can be achieved is to analyze the interactions between players and the agent, and use this information to direct the search towards behaviors that the players find interesting.

 

Ádám did his MSc in mechatronics at Budapest University of Technology and Economics. After graduation, he spent two years working on automated driving at Robert Bosch GmbH, during which he got exposed to both the classical and the machine learning approach of creating intelligent agents. His research interests include reinforcement learning, open-ended evolution and genetic encodings for neuroevolution.

Home institution: York

Supervisor: Professor Peter Cowling

Ready to apply?

Once you have identified your potential supervisor, we would encourage you to contact them to discuss your research proposal.

Learn More