Implementação e aplicação de algoritmos de aprendizado em um sistema neuro-simbólico

Abstract

One of the main goals of artificial intelligence is the creation of agents with humanlike intelligence. This has been researched using various approaches, and among the most prominent for machine learning are logic-based symbolic systems and artificial neural networks. Until the last decade, both approaches have progressed independently, but progress in both areas has led researchers to investigate ways to integrate both approaches. Several models that provide hybrid or integrated integration of these approaches emerged in the 1990s, and continue to be used to this day. This work has as main objective the implementation and use of the Neural-Symbolic conversion algorithm of Knowledge-Based Artificial Neural Networks (KBANN), the system has the ability to map a specific theoretical domain of rules (if-then) into a neural network, and refine the network using learning techniques. In addition, since the algorithm created by (Towell et al., 1990) does not have the capacity to acquire new knowledge and introduce them to the neural network, the algorithm TopGen (Optiz and Shavlik, 1995) will be used to add The network without losing the original knowledge acquired. The work used a board game to conduct experiments due to the well established rules of the game. The implemented system obtained interesting results, even with the initial rule domain perturbation (with the exclusion of them), obtaining an accuracy rate close to 100 %. Therefore, from the obtained results it was possible to conclude that the hybrid systems are able to overlap to adverse situations which were carried out the analyzes proposed in this research.