The Coevolution of Behavior and Motivated Action Selection

Abstract

The problem of action selection can be identified in ethology as the behavior switching problem. Here, a behavioral repertoire is chosen and then one behavioral module has to be selected until completion or its execution proves ineffective. In behavior based robotics this approach has been widely employed [Arkin (1998)], where an action selection mechanism (ASM) is used to arbitrate between several behavioral modules. The ASM helps in finding an immediate response to the basic question of intelligent systems of what to do next based on its previous sensory perception. However, it is important to notice that the right decision has to be made at the right time [Maes (1990)]. Therefore, some researchers in robotics consider important to develop a model able to decide the right time for making the right decision when solving a particular task. On the other hand, some other researchers are interested in providing a complete solution for the task to be solved. For example Nolfi proposes the evolution of a can-collection task using different motor neurons, and providing some sort of selector, but integrating the solution as a general behavior [Nolfi (1997)]. Potential problems related to the artificial separation of behavioral modules and to the process fusion of the behaviors support the development of general behavior. Also pointed out by the work of [Seth (1998). However, the works of [Kyong Joong & Sung Bae (2001); Yamauchi & Beer (1994)] offer an incremental solution for combining different architectures for solving an entire task. Here, the full integration of evolution with a central action selection model is the main interest. In this work, the behavioral modules are developed as separate components that can be integrated by the use of a selector. These modules can be implemented as neural networks, programming routines, or a mixture of both. The neural controllers may be optimized by the use of artificial evolution. An important feature of action selection is the emergence of opportunistic behavior [Brooks (1989)] that is not coded in the behavioral modules. We have based our current implementation on a model of central action selection that uses sensor fusion (CASSF) to build a uniform perception of the world in the form of perceptual variables [Montes Gonzalez, Marin Hernandez & Rios Figueroa (2006)]. However, in this implementation we added the use of two simulated motivations which correspond to fear and hunger. Also, the model is able to switch behaviors in normal and lesioned conditions (not explored in this paper). At this point we have decided to use staged evolution of behavior and then coevolution to adjust the weights of the neural controller for the selection mechanism. Hence, the evolutionary robotics approach was used in the design of three 23