Intelligent Agent to Negotiate on Goal Oriented Conversations
D.G.P.Y De Silva, P. Ravindra S. De Silva · 2020
In this study we propose a model to develop intelligent agents which are capable of negotiating on goal oriented conversations. These agents have the ability to learn negotiation using past experience and form strategies such as persuasion to negotiate successfully. In order to train these agents, they were made to interact with Simulated Users (SU). A corpus was generated and then annotated with speech acts to be used by a n-gram model. The SUs use this model to generate responses. In this study, we focused on single issue negotiations. A Markov Decision Process (MDP) was used to model the problem by defining the states and actions. The agent was made to interact with the SU to learn an optimal policy. For this we used the SARSA algorithm which is a temporal difference (TD) method. Once the agent is trained it was evaluated by a set of SUs built on different cultural norms. The number of dialogue turns and the policy score obtained when negotiating with these SUs were recorded and evaluated. It was observed that the agent was able to successfully negotiate to make a deal and also persuade them to a profitable offer. It can be concluded that the proposed model is successful and it can be used to train intelligent agents which can negotiate.