Learning by Doing vs. Learning from Others in a Principal-Agent Model∗

We introduce learning in a principal-agent model of stochastic output sharing under moral haz-ard. Without knowing the agents ’ preferences and the production technology, the principal tries to learn the optimal agency contract. We implement two learning paradigms — social (learning from others) and individual (learning by doing). We use a social evolutionary learning algorithm (SEL) to represent social learning. Within the individual learning paradigm, we investigate the performance of reinforcement learning (RL), experience-weighted attraction learning (EWA), and individual evolutionary learning (IEL). Our results show that learning in the principal-agent model is very difficult due to three main reasons: (1) the stochastic environment, (2) the discontinuity in the payoff space at the optimal contract caused by the binding participation constraint and (3) the incorrect evaluation of foregone payoffs in our sequential game principal-agent setting. The first two factors apply to all learning algorithms we study while the third is the main reason for EWA’s and IEL’s failures to adapt. We find that social learning, especially with a selective replication operator, is much more successful in adapting to the optimal contract than the canonical versions of individual learning from the literature. A modified version of the IEL algorithm using realized payoffs evaluation performs better than the other individual learning models; however, it still falls short of the social learning’s performance