Automatic acquisition of dialog strategies for concept learning through interaction among agents

This paper describes automatic acquisition of dialog strategies for efficient concept acquisition applied to the interaction among infant agents (IAs). The following two topics are emphasized: (1) what the influence of dialog strategies over concept acquisition is, (2) how cooperative dialog strategies are acquired. In order to acquire efficient strategies, Q-learning is applied to a strategy acquisition mechanism. Experiments are carried out through the interaction between two IAs; one is teaching IA (T-IA) that has all the concepts and teaches them to the other IA, and the other is learning IA (L-IA) that acquires concepts. In the experiments, 100,000 times of dialogs arc iterated and both IAs acquire dialog strategies. We first fix the number of initial concept of the L-IA to zero and compare the acquired strategies with a base line strategy of Random-Strategy, which makes the act of TA randomly, to investigate the above-mentioned topics of (1) and (2). The experimental results showed that the resultant strategy, which follows Grice's maxim of "quantity", makes the concept acquisition efficient, however, the T-IA using this strategy exhibits the same behavior no matter how the L-IA asks. This is not the cooperative behavior. In the second experiment, zero to ten initial concepts are randomly given to the L-IA for setting actual dialog situation. The results showed that both IAs acquire cooperative strategies that can efficiently eliminate ambiguities. The behavior of IAs using the cooperative strategies is similar to the one that is seen in human-human interaction