Learning in complex environments with feature-based categorization
Alejandro Agostini, Enric Celaya · DIGITAL.CSIC (Spanish National Research Council (CSIC)) · 2005
The application of reinforcement learning techniques in the typically complex environments of the autonomous agents is only feasible when some kind of generalization among situations is possible in order to reduce the number of experiences required for the learning task. Many reinforcement learning techniques have been proposed to cope with this problem but they usually do not take advantage of all the opportunities of generalization. A new algorithm was proposed in [12] that exploits a type of regularity that is denoted as categorizability. Categorizability means that from all the relevant features that must be taken into account to decide the best action in any situation, only a few of them are actually relevant in each particular situation. In this paper the categorization and learning capabilities of the algorithm are evaluated using a problem which satisfies to a good extent the categorizability property. The categorization achieved by the algorithm in this problem is analysed in detail and illustrated with examples. The learning performance of the algorithm is compared with those of other reinforcement learning algorithms. Some improvements of the original algorithm are introduced.