Neural Network Based Approach for Learning Planning Action Models

Abstract

Artificial Intelligence (AI) planning technology represents one of the most efficient solutions to guide goal driven agents in domain problems which are mostly characterized by problem solving features. The current proliferation of physical and virtual autonomous agents have therefore triggered a growing interest in the automatic acquisition of action models, which can take advantage of the observation capabilities of agents, which sense world states through their sensors. In this work we present a Neural Network (NN) based approach for learning AI planning action models by observation in noisy environments. The system learns by observing a set of execution patterns of the same action in different contexts. The perceptions of both pre/post action execution states are used to train the NN learning component and are assumed to be affected by noise, which could be due to inaccurate reading or malfunctioning of the sensors. Preliminary experiments shows that the proposed NN learning module seems to be more resilient to unforeseen situations with respect to traditional propositional-based approaches to action model learning.