First published 2 September 2013
Evolving Behaviour-Dependent Strategies in Agent Negotiations
Darius Falahat, Enrico Gerding, Markus Brede
We use genetic algorithms to evolve trading strategies for iterative bilateral negotiations between buyers and sellers. In contrast to previous work we evolve purely reactive strategies that base decisions on memories of behaviour in previous negotiation rounds. We find that simulations lead to three main types of behaviour: (i) cooperative outcomes in which bargaining leads to an agreement and equal sharing of profits, (ii) uncooperative outcomes in which negotiations are not successful and (iii) outcomes in which one party profits at the expense of the other. The frequencies of each type of behaviour vary when the probability for negotiations to terminate is changed, confirming our hypothesis that cooperation should decrease as this break-off probability increases. Comparisons of the results to tit-for-tat (TFT) strategies and previous research on the iterated prisoner's dilemma (IPD) are used to understand simulation results, and we observe the emergence of TFT behaviour during periods of agent cooperation.