Skip to main content

Learning from Failure

Staff Report 299 | Published February 1, 2002

Download PDF


Learning from Failure


We study decentralized learning in organizations. Decentralization is captured through a symmetry constraint on agents’ strategies. Among such _attainable strategies_, we solve for optimal and equilibrium strategies. We model the organization as a repeated game with imperfectly observable actions. A fixed but unknown subset of action profiles are _successes_ and all other action profiles are _failures_. The game is played until either there is a success or the time horizon is reached. For any time horizon, including infinity, we demonstrate existence of optimal attainable strategies and show that they are Nash equilibria. For some time horizons, we can solve explicitly for the optimal attainable strategies and show uniqueness. The solution connects the learning behavior of agents to the fundamentals that characterize the organization: _Agents in the organization respond more slowly to failure as the future becomes more important, the size of the organization increases and the probability of success decreases_.