Grid-Robot Drivers: an Evolutionary Multi-agent Virtual Robotics Task

Abstract

Beginning with artificial ants and including such tasks as Tartarus, software agents that are situated on a grid have been a staple of evolutionary computation. This manuscript introduces a grid-robot problem in which the agents simulate single or multiple drivers on a two-lane interstate freeway that may have obstructions. The drivers are represented as if-skip-action lists, a linear genetic programming structure. With one driver present, the problem is similar to an artificial ant task, requiring only that the grid robot learn where fixed obstacles are placed. When multiple drivers are present, the process of driving can be cast as a game similar to the prisoner's dilemma. The relative advantage to be gained from inducing another vehicle to crash is analogous to defection in the prisoner's dilemma. The game differs from prisoner's dilemma in that defecting is a complex learned behavior, not simply a move the grid robot may choose. A skilled opponent may dodge an attempt at "defection". Six sets of experiments with up to five drivers and two fixed obstacles are performed in this study. In multi-driver simulations evolution locates a diversity of behaviors within the context of the driver task