Neuronal networks working at multiple temporal scales as a basis for amphibia’s prey-catching behavior

We analyze a model of neuronal mechanisms underlying amphibia's prey-catching behavior, integrating hypotheses generated within different areas of Neuroscience and studying how the efficacy of visual prey-like dummies to release toad's prey-catching actions depends on parallel distributed processes occurring at multiple levels of temporal abstraction. First, in the scale of 100's of msecs, changes in neuronal activity caused by the stimulus characteristics and its current spatial-temporal relationship with the toad, as well as nervous signals related to actions' expected consequences (e.g., mouth mechanoreceptors activation after a snapping); second, signals generated during learning events happening at a temporal scale of minutes to hours; third, signals related to the course of actions, within an undetermined time scale that may last for several hours; and fourth, signals generated by changes in motivational factors (e.g., hunger, daily and yearly cycles) occurring at a much slower time scale. In addition, we analyze how in this knowledge representation, the course of actions (plan) is episodic, goal-oriented and can be modulated by learning, or by changes in the agent's motivational state. This modulation is the outcome of accommodating information of new situations (a non catchable prey-like stimulus) into the dynamics of underlying neuronal mechanisms, in order to change the way the toad (agent) normally responds to that domain of interaction (stop yielding prey-catching behaviors towards that specific stimulus), without affecting its performance when similar situations appear in its immediate surroundings (prey-catching behaviors to real prey remain unchanged). Knowledge representation at different levels of temporal abstraction, and its relationship to planing and learning are key issues in Distributed Artificial Intelligence, and in behavior-based robotics (Sutton and Barto, 1998; Beer et al, 1990). In particular when the aim is to build agents that could incorporate (" learn ") into their control (planing) system different aspects (" knowledge ") of the relationship they maintain with other agents or with an uncertain environment, i.e., their " ecology " (Arkin et al, 2000). Some work in the field of behavior-based robotics have shown the benefits of exploiting models of behavior proposed by In our previous work (Arkin et al., 2000), we used Schema Theory as an interlingua to implement a model of the agent-environment interactions of the Praying Mantis (Cervantes-Pérez et al, 1993a) into the control system of a robotic hexapod, Hermes II, that displays visually guided behaviors, such as prey-acquisition, mating, predator-avoidance, obstacle-avoidance, and the " chantlitaxia " behavior. Continuing our …