Learning Policies for Blocks

A. González Romero, A. Gonzales Camargo, and R. Alquèzar (Spain)


Planning, policies learning, evolutionary algorithms, blocks world.


The principal aim of this article is to investigate whether good automatically learned policies can be generated using training examples along with an Evolutionary Algorithm that uses a certain method (Algorithm 1) to calculate the fitness of a policy. The training examples are used as input to the learning algorithm; they describe a number of solved instances in a certain domain. Our work deals with the blocks world domain. A system has been developed where a set of policies is randomly generated and then evolved. Each policy is composed of an ordered set of rules; a fitness value is assigned to each policy according to its performance on the training examples using Algorithm 1. A policy that solves any blocks world problem, though not optimal, has been learned in this way.

Important Links:

Go Back