Learning to manage risks in non-cooperative dialogues

We investigate 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. Here we present new results showing how learned non-cooperative dialogue strategies change depending on a) how severe the penalty is for being caught being non-cooperative, and b) how risky the non-cooperative behaviour 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 an additional 10% chance of being caught (exposed) every time it attempts a non-cooperative (manipulative) move, when the penalty for being caught is that the adversary will no longer trade.