Grant: Applied Research and Development Grants - Level 1

Reinforcement learning is learning what to do - how to map situations to actions - so as to maximize a numerical reward signal. The learner is not told which actions to take, but instead must discover which actions yield the most reward by trying them. Since RL requires a lot of data, therefore it is most applicable where simulated data is readily available like gameplay, robotics. Here, we consider specifically systems that interact with users in the loop. When a human is involved, there are recurring issues: there is rarely a good simulator, the systems are stochastic and non-stationary, and running them is expensive and/or slow. This is very different from training on a simulated environment where data is effectively unlimited, consequences for poor actions are non-existent and system dynamics are clean and often deterministic. An open-source framework for such a training environment is proposed by the company AI-R, and aims at solving problems of training reinforcement learning algorithms with a human in the loop: non-rewindability, delayed reward functions and unbalanced relevance. The goal of this research project is to use their framework to train a learning agent to teach a human., and we will try to demonstrate that we can effectively teach a human player to play the game Smash Bros. This would enable academics in various fields to use this framework for research where a human behaviour is studied, and open the door to a range of new products and business models for entrepreneurs to develop.