Collaboration among computational agents through indirect reciprocity

Abstract

Indirect reciprocity is a mechanism for collaboration from evolutionary theory that attempts to explain the occurrence of altruism. It explains how a population of entities can make a decision of whether to cooperate in an interaction on the basis of the reputation of others. This basis of collaboration is largely seen in humans. We possess superior communicative and cognitive abilities compared to other species, enabling us to share and evaluate information through rumors of what happens between other people. There are limitation in the effectiveness in the indirect reciprocity we see in humans. People only collaborate when they have a reason to do so, and will hold back if there is uncertainty if an act of helping will ever be reciprocated. The lack of trust and uncertainty increases with the number of people that partake in such a mechanism, as we have a hard time communicating and reasoning about large amounts of information. We identify that uncertainty in the evaluation of other peoples reputation, as well as the cost of maintaining a reputation on an entire population acts as limitations of indirect reciprocity. Representing humans through computational agents allows us to increase the mere size of information we are able to communicate and reason about. We provide agents with a public observation model using blockchain technology and smart contracts, enabling them to have access to information about encounters between the other agents. The agents are granted with certainty that the information from the observation model is correct. They sign contracts as proof that interactions occur using asymmetric encryption. The resulting contracts are used as a basis for the evaluation of reputation of other agents. We use social norms to define good behaviour, and consequently how the agents assess the reputation of other agents. Every agent conforms to a social norm of their choice, and as a consequence they have a subjective opinion on which agents have good reputation. Through social learning by imitating agents that perform better, the population converges to a homogeneous one in terms of social norms. Using a strict discriminating social norm such as stern-judging, we see a near perfect cooperation rate where the population successfully discriminates agents employing other social norms.