Can an agent learn how to deceive its interlocutor?

Non-cooperative dialogue behavior is important in a variety of application areas, including education, military operations, video games and healthcare. However, it has not been addressed using statistical approaches to dialogue management, which have always been trained for co-operative dialogue. We develop and evaluate statistical dialogue agents which learn to perform non-cooperative dialogue moves in order to complete their own objectives in a stochastic trading game. We show that, when given the ability to perform both cooperative and non-cooperative dialogue moves, such an agent can learn to bluff and to lie so as to win games more often -- against a variety of adversaries, and under various conditions such as risking penalties for being caught  in deception. We also show how learned non-cooperative dialogue strategies successfully change depending on a) how severe the penalty is for being caught being non-cooperative, and b) how risky the non-cooperative behavior is (i.e. the probability of being caught). For example, we show that a  non-cooperative dialogue agent can learn to win an additional 4.5% of games against a strong rule-based adversary, even when there is a 10% additional chance of being caught every time it attempts a non-cooperative move, if the penalty for being caught is that the adversary will no longer trade.