Evolution of social learning strategies

We study three types of learning with Bayesian agent-based modeling. First, we show that previous results obtained from learning chains can be generalized to a more realistic lattice world involving multiple social interactions. Learning based on the passing of posterior probabilities converges to the truth more quickly and reliably than does learning based on imitation and sampling from the environment; and the latter method gets closer to the truth than does pure imitation. The passing of posterior probability distributions can be viewed as teaching by explanation, and as an implementation of the cultural ratchet, which allows rapid progress without backsliding. We also show that evolution selects these learning strategies in proportion to their success. However, if the environment changes very rapidly, evolution favors the imitation-plus-reinforcement strategy over the more sophisticated posterior passing. Implications for developmental robotics, human uniqueness, and interactions between learning and evolution are discussed.