Self-motivated learning of achievement and maintenance tasks for non-player characters in computer games

This paper presents a framework for motivated reinforcement learning agents that can identify and solve either achievement or maintenance tasks. To evaluate and compare agents using these approaches, we also introduce two new metrics to better characterise and differentiate the behaviour of characters motivated to learn different kinds of tasks. These metrics quantify the focus of attention and dwell time of agents. We perform an empirical evaluation of motivated reinforcement learning agents controlling characters in a simulated game scenario, comparing the effect of three different motivations for learning achievement and maintenance tasks. Results show that we can generate characters with quantifiably different achievement and maintenance oriented behaviour using our proposed task identification approach. Of the three motivations studied - novelty, interest and competence - novelty-seeking motivation is the most effective for creating agents with distinctive maintenance or achievement oriented behaviours.