Evolving Behavior Coordination for Mobile Robots Using Distributed Finite-State Automata [chapter]

Pavel Petrovic
2008 Frontiers in Evolutionary Robotics  
This chapter describes a finite-state automata a p p r o a c h t o E v o l u t i o n a r y R o b o t i c s ( E R ) (Petrovic, 2007) . Most of the efforts in the field of ER, see (Nolfi and Floreano, 2000) for very representative examples concentrate around controllers with various neural architectures, mainly due to their high plasticity and adaptability. Despite these obvious advantages, the resulting controllers usually have monolithic form with information represented numerically across the
more » ... hole networks, where it is difficult to understand and explain why particular actions are taken in particular situations, enumerate the possible types of behaviors the controllers may produce, and achieve modularity required in more complex tasks. These features may be needed when the field reaches the level of real-world applications, for example to automatically verify the safety. In addition, many robotics researchers have demonstrated and reasoned that modular controller architectures with simultaneously executing behavioral modules -often called Behavior-Based (BB) architectures, are favorable as compared to centralized and top-down architectures. In our quarter of the ER community, the ultimate goal of the efforts is to find useful methods for automatic programming of mobile robots performing non-trivial tasks. By non-trivial, we mean tasks for which manual controller design by an engineer is difficult, or applications where the engineer is not available at the time of controller design or adaptation. Modular neural networks are an interesting field of interest, however, the research results did not satisfy us in their ability to compensate for their shortcomings. We suggest to use a BB architecture and apply Evolutionary Computation (EC) to learn/design the coordination mechanism of the behaviors. Although it would be possible to use EC to evolve both the behaviors and the coordination, evolving simple behaviors have been addressed by many previous works, and thus this time, we concentrate at the coordination mechanisms of a set of pre-programmed (pre-evolved) behaviors. When searching for a suitable formalism for representing the coordination mechanisms, we found that the robot behavior can be veritably modeled by finite-state automata (FSA). We also found that EC have been previously successfully applied to automatic programming of state automata, a detailed overview is in the section 3. As explained in section 4, we have designed a set of experiments, which indicate that for tasks that share properties with BB coordination mechanisms, the finite state representations outperform more conventional GP-tree representations. With this background, we have performed experiments with evolving
doi:10.5772/5467 fatcat:7hmgxgyrevfb7d4vsmbjesep5y