Predictive automated negotiators with different attitudes towards risk

摘要

A predictive strategy employed by an autonomous agent who engages in multi-issued negotiations is presented and assessed. The strategy allows the predictive agent to adapt his consequent offers with respect to the estimations of his counterpart's responses. As different attitudes towards risk may emerge, a risk-related parameter is embodied to the strategy. Forecasting is conducted with the use of Multilayer Perceptrons (MLPs) and the training set is extracted during the negotiation session. Two cases are examined: one where separate MLPs are used to estimate each negotiable attribute and one where a single MLP is used to estimate the counterpart's response. It is shown that simple MLPs with one hidden layer are adequate for forecasting the counterpart's offer vectors, and are tested in order to find the appropriate number of nodes on input and hidden layer. The network that yields the lowest error, incurs to the predictive agent 10.5% average absolute increase of his individual utility (gain).