Abstract
This paper proposes an artificial player for the Turtle Race game, with the goal of creating an opponent that will provide some amount of challenge to a human player. Turtle Race is a game of imperfect information, where the players know which one of the game pieces is theirs, but do not know which ones belong to the other players and which ones are neutral. Moreover, movement of the pieces is determined by cards randomly drawn from a deck. The artificial player is based oil a non-linear neural network whose training is performed by means of a novel parallel evolutionary algorithm with fitness diversity adaptation. The algorithm handles, in parallel, several populations which cooperate with each other 133, exchanging individuals when a population registers a diversity loss. Four Popular evolutionary algorithms have been tested for the proposed parallel framework. Numerical results show that an evolution strategy can be very efficient for the problem under examination and that the proposed adaptation tends to improve upon the algorithmic performance Without any addition in computational overhead. The resulting artificial player displayed a high performance against other artificial players and a challenging behavior for expert human players.