Biologically-motivated neural learning in situated systems

摘要

We describe the autonomous robot ARBIB. This uses biologically-motivated forms of learning to adapt to its environment. ARBIB'S 'nervous system' has a non-homogeneous population of spiking neurons, and uses both nonassociative and associative forms of learning to modify pre-existing ('hard-wired') reflexes. As a result of interaction with its environment, interesting and 'intelligent' light-seeking and collision-avoidance behaviors emerge which were not pre-programmed into the robot (or 'animat'). These behaviors are similar to those described by other workers who have generally used behaviorally-motivated reinforcement learning rather than biologically-based associative learning. The complexity of observed behavior is remarkable given the extreme simplicity of ARBIB's 'nervous system', having just 33 neurons. We take this to indicate that great potential exists to explore further "the animat path to AI".