On using Brainwaves as Implicit Human Feedback in Reinforcement Learning

In this work, we explore an interesting solution paradigm that allows humans to assist machine learning algorithms. Specifically, we study the use of electroencephalogram (EEG) based brain waves of the human to generate implicit feedback that are then be used by the Reinforcement Learning (RL) algorithms. Such a model benefits from the natural rich activity of the human brain, but at the same time does not burden the human observer since the feedback being relied upon is intrinsically generated. This paradigm is based on a high-level error-processing system in humans that naturally generates an error-related event potential (ErrP) when the observer recognizes an error made by the agent. In order to systematically study different aspects of the paradigm, we use computer games as proxies for the real-life environments that the agents will need to operate in. The use of games as proxies for real-life has some distinct advantages including a highly controllable and replicable environment.