Automated weighting of environment features in combat-based games

Abstract:

A major goal of games research is to engineer artificial systems to play at or above human skill levels. Intuitively, in many ‘shooter’-like games, an agent might move based on features of the environment; for instance, an agent might reposition to areas of cover or with strong line-of-sight. While many possible desirable features exist, it is often unclear how to asses their relative value to guide agent behavior. One technique for determining the relative value of parameters is called Particle Swarm Optimization (PSO), in which a collection of agents with different weightings gradually converge towards a solution. We developed a simple grid-based, turn-taking shooter environment and automated the assessment of each location in terms of various desirable features. We then generated agents with random weightings for the value of these features and performed PSO over a series of combats between them. We assessed two types of combats: static fights, in which each agent may freely choose its starting location but cannot subsequently move, and dynamic fights, in which agents may reposition relative to each other mid-conflict. Despite the stochastic nature of this approach, results suggest the existence of a unique best strategy for static fights under our assumptions.