Believable judge bot that learns to select tactics and judge opponents

This paper describes our believable judge bot ICE-CIG2011 that has an ability to learn tactics from a judge player and an ability to judge an opponent character as a human or a bot. We conjecture that a bot with these two abilities should be considered human-like in a competition environment, such as BotPrize, where human players participate to compete not only for being the most human-like player but also the best judge. Main contributions of this work lie in our mechanisms for achieving these two abilities. To achieve the former ability, we develop a system and GUI that allow a selected judge player - whose role is to train ICE-CIG2011 - to control his or her character by only deciding which tactic to use under a given situation. We then obtain the judge´s tactic log and use it for training tactic selection of ICE-CIG2011 with neuro evolution of augmenting topologies. To achieve the latter ability, we acquire additional logs when the judge character interacts with other opponent characters. In order to represent the play of a known (bot or human) character, we train a neural gas - a kind of self-organizing neural network - from its log. For an unknown character, once its neural gas is trained after a certain period of observation, ICE-CIG2011 decides if it is a human or bot by using the K-nearest-neighbor algorithm; this algorithm considers the majority in the labels of the K-nearest neural gases, of known characters, to the neural gas of that unknown character. Experimental results are given and discussed concerning these two abilities of ICE-CIG2011.