Probabilistic Learning Strategies Applied to Agreement Negotiation for Meeting Scheduling -Edición Única

Abstract

This document presents the thesis required to get the degree of Master in Science in Intelligent Systems. One of the problems more commonly found in Artificial Intelligence is related to negotiation where all participants involved in the decision process have to reach an agreement in a way that minimize costs for each participant; this problem is known as Distributed Agreement Problem. Our specific problem is focused on Multiagent Meeting Scheduling where several variables are considered for selecting the best schedule which fulfills the requirements and preferences of each participant. Recently there is a necessity to find optimal solutions to problems of this type to satisfy the demands on time, quality and productivity into a competitive society. The research work described in this thesis is mainly focused on probabilistic multiagent learning for the selection of the best strategy applied in specific scenarios. We are using the approach of Crawford ([Crawford 05b]) about Playbooks to select the best negotiation strategy taken for an agent that acts as organizer or invitee in a meeting scheduling negotiation. In this approach the agent who uses the playbook is the only learning agent in the environment, the others agents are using static strategies. In our research work we implement several negotiation strategies and use the playbook to select the best one; our contribution is to use new and adapted strategies using several heuristics to make the agents more cooperative or competitive. This research work is focus on the development of modified algorithms applied to our multiagent meeting scheduling, in specific modified negotiation strategies helped by some learning algorithms. We put all these together and did a set of experiments to see the complexity, usefulness and performance of the algorithms in several scenarios