Beyond Playing to Win: Creating a Team of Agents With Distinct Behaviors for Automated Gameplay

In this article, we present an approach to generate a
team
of general video game playing agents with differentiated behaviors that can ultimately assist in the game development process. We consider the agent behavior as the corresponding outcomes of playing the game: rate of wins, score, exploration, enemies killed, items collected, etc. We create and identify agents that are expected to achieve particular goals but do not necessarily simulate human behavior during gameplay. We present a solution that, by
heuristic diversification
, provides a controller with different heuristics and a corresponding set of
weights
, driving its actions. Given the simplicity of this
behavior-encoding
and its easiness to evolve, we use Multidimensional Archive of Phenotypic Elites to generate different solutions that elicit particular behaviors and assemble a
team
. The resulting agents are allocated in a feature space, used to identify the expectations of each of them. We generate a
team
for four games of the General Video Game Artificial Intelligence framework and find six different
behavior-type
agents in each. We include an experiment to check the portability of these agents when playing alternative levels and an exploratory work aiming to use them to detect design flaws in game levels.