SESiL: Social, Evolutionary Supported Learning

Social Learning describes several variations of interaction between a learning agent and a source of (potentially) beneficial behaviour. Mainly, though outliers exist, there are three forms of this interaction. First is the obvious “monkey-see-monkey-do”, learning by the imitation of an observed behaviour. Second, the “teacher-learner” relationship, where an experienced agent actively guides or instructs the learner. Finally, knowledge extraction from observation, whereanagentgeneralizes from theobservedinteractionsof others with the environment to its own needs and goals. However, in spite of the enormous volume of work in social learning, once commonality persists — an agent can directly benefit from the additional information and improve its own behaviour. But what happens if the agent has identified the benefits of other’s behaviour, but cannot absorb them? In this paper, we study the support that social learning can garnerfromanevolutionaryperspectiveontheprocess: ratherthan absorbing additional behavioural information directly, agents share and merge their behavioural information by choosing a mate. It is the children that represent and carry the socially learned, combined behaviour. We term the combined learning process SESiL (Social, Evolutionary Supported Learning). Besides the formal definition of the framework, we provide experimental studies of its properties. Specifically, wedeploySESiLinmulti-taskingclassification. Starting from a population of agents who have been partially-pretrained on small subsets of labels, we give them the agency to seek and choose a mate based on the observed classification performance. Presuming availability of a “genetic merger” operator (in our case, classifier network merger), we allow the mutually-agreed mating pairs to be replaced by two children that carry their (imperfectly) merged knowledge. We baseline SESiL against a full-data access classifier, a distributedlearner(split-learn-merge) andseveralforms ofmoreclassicalevolutionarycompute, whereagentshavenosayin choosing a mate, but are bred following their overall performance.

Authors

• Tianshu Zhao Carleton University
• Zinovi Rabinovich Carleton University

Keywords

• Multi-Agent Learning
• Social Learning
• Evolutionary Computing