Final Report Summary - CAP (Computers Arguing with People)
The goal of the CAP project was to develop models and algorithms that enable automated systems to argue proficiently when interacting with people. The arguer agent would like the outcome of the negotiation to unfold according to its own preferences. There are several ways to argue with and persuade people, and we considered several approaches and developed and tested several algorithms for various aspects of negotiations. Our goal was to enable the agents to negotiate in more natural ways (e.g. using natural language rather than menu driven) and reach better agreements according to their specifications. We studied both game theory-based methods and knowledge-based methods. The development of algorithms and techniques to predict people’s decision-making played an important role in the development of our agents. The prediction techniques deployed facial expression, humans’ actions or their messages. The challenge was providing high accuracy predictions using a relatively small and unbalanced number of training examples.
Notable agents that we developed include NegoChat agent, the first negotiation agent to successfully negotiate with people in natural language; the Personality Adaptive Learning (PAL) agent which negotiates with people from different cultures; SAP, a Social agent for Advice Provision which generates advice according to a social model that we developed; equilibrium agents which follow strategies that are in equilibrium; the Sigmoid Acceptance Learning Agent (SIGAL), which uses a decision-theoretic approach to negotiate in revelation games, which is based on a model of how humans make decisions in the game; and the DIG agent that can assist in detecting and incriminating a deceptive participant in a chat-room. The SPA agent is the first agent to present humans with arguments in a dialog. It combines theoretical argumentation modeling, machine learning and Markovian optimization techniques. The Virtual Suspect (VS) is able to play the role of a suspect in simulations used for training young law-enforcement personnel in interrogative interviews.
Notable agents that we developed include NegoChat agent, the first negotiation agent to successfully negotiate with people in natural language; the Personality Adaptive Learning (PAL) agent which negotiates with people from different cultures; SAP, a Social agent for Advice Provision which generates advice according to a social model that we developed; equilibrium agents which follow strategies that are in equilibrium; the Sigmoid Acceptance Learning Agent (SIGAL), which uses a decision-theoretic approach to negotiate in revelation games, which is based on a model of how humans make decisions in the game; and the DIG agent that can assist in detecting and incriminating a deceptive participant in a chat-room. The SPA agent is the first agent to present humans with arguments in a dialog. It combines theoretical argumentation modeling, machine learning and Markovian optimization techniques. The Virtual Suspect (VS) is able to play the role of a suspect in simulations used for training young law-enforcement personnel in interrogative interviews.