Predictive decision-making mechanisms based on off-line and on-line reasoning

Negotiation is an interaction allowing for solving con°icts between multiple parties. The process of negotiation involves different types of decisions. Before the interaction starts the agent needs to choose its partner, and during the interaction the agent needs to decide on what ffers to propose and what offers to accept. The outcome of these decisions may be improved when the agent uses forecasting, through the learning and reasoning techniques, of prospective partners' behaviour. In this thesis two predictive decision-making mechanisms are proposed. The first mechanism allows for the selection of negotiation partners, and the second allows for the generation of negotiation offers during the actual negotiation. The selection mechanism uses the memory of previous encounters to make prediction about the potential behaviour of prospective negotiation partner. The prediction is obtained in a form of probability or possibility distribution for each potential negotiation partner, assigning to each potential negotiation outcome a degree of belief that this outcome will occur in the current negotiation. The distribution corresponding to the modelled agent is then fused with the utility function of the decision-maker which yields the value of expected utility. The agents maximizing the value of expected utility are considered the best prospective partners. The second decision-making mechanism uses the sequence of offers of both negotiating parties, up to the current negotiation stage, to predict the concession curve of the opponent. This prediction takes the form of a function mapping potential future offers of the modelling agent into the next offers of its counterpart. The construction of the concession curve forecast employs the regression analysis. The agent fits the memory of offers to each of the possible models of negotiation behaviour. The model that fits best to the data is then chosen for the forecasting. The forecast obtained in such a manner is used to generate offers during the remaining negotiation rounds. The most important contributions of this dissertation to the subject matter are the new application of the concept of case-based reasoning, the adaptation of the case-based reasoning mechanism to the negotiation context, the new algorithm for the estimation of the possibilistic expected utility, the application of the regression analysis to the forecasting of the multi-tactic negotiation strategy and the discussion of the limitations of the regression analysis. Both decision-making mechanisms are validated experimentally.