Birds in Boots: Learning to Play Angry Birds with Policy-Guided Search

Abstract

In this paper we present Birds in Boots (BiB), a system that uses a sampling-based search algorithm to learn a neural policy for solving Angry Birds levels. Our learning procedure is based on the Bootstrap algorithm, which was previously used to learn heuristic functions for solving classic heuristic search problems. BiB starts its learning procedure with a policy given by a randomly initialized neural network. This initial policy is used to guide the search algorithm on a set of procedurally generated Angry Birds levels. The levels the search algorithm is able to solve are used to improve the neural policy. We repeat this procedure a number of times, until solving all levels or reaching a time limit. We perform several experiments with different instances of our method and show that it can solve more levels than other approaches, including learning-based and rule-based methods.