Evolving Neural Network Agents in the NERO Video Game

摘要: In most modern video games, character be- havior is scripted; no matter how many times the player exploits a weakness, that weakness never repaired. Yet if game characters could learn through interacting with player, behavior improve during game- play, keeping it interesting. This paper introduces real-time NeuroEvolution of Augmenting Topologies (rt- NEAT) method for evolving increasingly complex arti- ficial neural networks in real time, as being played. The rtNEAT allows agents to change and game. fact, makes possible new genre games which teaches team series customized training exercises. order demonstrate this concept NeuroEvolving Robotic Operatives (NERO) game, trains robots combat. describes results from novel application machine learning, also demonstrates multiple can evolve adapt like NERO time using rtNEAT. future, may allow kinds educational applications online user gains skills.