Towards Motivated Reinforcement Learning

Abstract

Imbuing artificial agents with the ability to learn allows them to change their structure and improve their performance at a task [1]. Learning agents have the potential to make computer games more interesting because they can adapt in ways that reflexive agents cannot. Supervised learning, for example, has been used for behavioural cloning of players in multiuser games. The supervised learning agent observes a player’s actions and learns to represent the player’s avatar when the player is not online. Reinforcement learning [2] has been used to create non-player characters that can adapt their behaviour for specific tasks – such as fighting – in response to their opponent’s behaviour during game play [3]. The focus of this book, motivated reinforcement learning, allows a new kind of non-player character that can not only adapt its behaviour for individual tasks, but also autonomously select which tasks to learn and do.