MAP-Elites to Generate a Team of Agents that Elicits Diverse Automated Gameplay
Cristina Guerrero-Romero, Diego Pérez-Liébana · 2021 IEEE Conference on Games (CoG) · 2021
The objective of this work is to provide a procedure to generate a team of players for a game so they are available to the developer to choose from and that can be used for automated gameplay. Our solution applies the MAP-Elites algorithm to generate agents with distinct behaviours. The resulting agents are distributed in the space of features based on the result of their actions when playing a game: wins, score, % explored, interactions, kills, items collected, etc. The criteria used as the performance of the elites does not come from how well an agent plays a game, but by the time it takes it to play it and determine the cell in the map it falls into. We present and implement the solution and include details about the agent used, as well as the list of heuristics created to elicit differentiated behaviours within the game, representing distinct types of players. We executed the algorithm implemented for three different games and a total of 33 configurations. The size and diversity of the pool of agents generated allow running automated gameplays in each of the games to elicit different expected behaviours. The options are limited by the distribution of the team within the space, given by the pair of features or the characteristics of the game. The methodology gives the flexibility to extend the features or modify the range of existing ones to have control over the behavioural space and, therefore, the characteristics of the generated team.