Autonomous robotic reactive behavior: Synthesis, scalability, and transparency

Elements of autonomous robotic reactive behavior synthesis, scalability, and transparency are investigated in this thesis. Artificial neural networks (ANNs) are used to synthesize the behavior of reactive wall following. The goal is to achieve wall follow behavior generalization, which is tested across numerous varying environments. Behavior scalability is examined by scaling the operating speed of the trained ANN to more challenging speeds and comparing the resulting wall follow trajectory performance to the human operator counterpart. Despite their successes, ANNs succumb to the problem of poor transparency. Behavior transparency is desirable so that proper operation can be verified by a human expert. Therefore, the problem of enhancing ANN behavior transparency is explored by extracting comprehensive rule sets from the trained ANN. The proposed method consists of discretization, feature selection, and evolutionary rule-set search using a real-coded genetic algorithm. Experimentation is conducted on three different ANN architectures with varying size and complexity.