Artificial Theory of Mind in contextual automated negotiations within peer-to-peer markets

Abstract

With the advent of the sharing economy (eBay, Airbnb, Uber) and peer-to-peer electricity markets, negotiation agents become key to automate trading between prosumers. To be competitive, agents need to develop models of their opponents? strategic behavior depending on contextual circumstances. However, existent approaches do not account for others? ways of reasoning about how the context influences opponents? strategies. In this work, negotiation agents are endowed with artificial Theory of Mind (aToM) to represent mental states of others. Those states include their strategies, beliefs, the context influencing beliefs and the beliefs they may have about the beliefs others have about themselves depending on the context. From prior knowledge that relates private and contextual states with opponent´s plausible strategies, agents build models of different orders of aToM using weighted Gaussian Processes with heteroscedastic noise and adapt their strategies. During a negotiation episode, each agent selects the order of aToM that best accounts for its opponent´s counteroffers based on Bayesian surprise. To assess the proposed approach, a benchmark Smart Grid is used, where energy prosumers negotiate on a peer-to-peer basis to compensate for their energy deficits and sell their surpluses. Results obtained highlight that low orders of aToM increase social welfare at least in a 54% despite the selfish nature of prosumers. Also, more sophisticated aToM abilities increase the individual competitiveness of agents. This is the first work in the related literature that adds aToM abilities to negotiation agents while considering the negotiation context.