Basal Ganglia Based Robot Behavior Selection And Behavior Sequence Learning Methods Study

This dissertation studies the basal ganglia inspired robot behavior selection and behavior sequences learning methods, as well as FPGA implementation techniques of the basal ganglia. The main parts are concluded as follows:Two behavior selection mathematic models of the basal ganglia are analyzed and compared, and their parameters are optimized by GA. Selection abilities before and after the optimization are then compared through simulation experiments. Selection model of the basal ganglia is then embedded in behavior-based control architecture and its selection ability is demonstrated by a simulated robot food collection task. Two typical movement disorder diseases caused by dysfunction of the basal ganglia are studied in detail through simulations. Behavior patterns of a real robot when it "gets sick" are also studied.Properties of spike neuron model and synapse plasticity are studied in detail. The information processing similarity between fuzzy control and spike neuron model are analyzed. A simulated "eye-arm coordination" experiment is then carried out to study the reciprocal relationship between two nuclei. Then relationships between the multiple channels and nuclei within the basal ganglia are then studied. Spiking neural network of the basal ganglia is constructed and its selection ability is demonstrated by simulation and real robot experiments.Dopamine activity property during condition response, as well as its modulation effect on synapse plasticity is studied. And potential applications of this kind of modulation effect are explored through simulated maze robot and character recognition experiments. The underlying biology substrate of the reinforcement learning mechanism of the basal ganglia is the analyzed before the Actor-Critic reinforcement learning architecture is proposed. The effect of the basal ganglia in behavior sequence learning is demonstrated by the simulated spelling robot.Spike neuron model is transformed to make it more conveniently implemented by digital logic circuits. FPGA implements of single spike neuron model and synapse are written in Verilog HDL program. The total architecture is first designed and the functions of sub-modules are specified. All the sub-modules are simulated, synthesized and implemented before combined together to form the top module. Finally, basal ganglia-one of the important structures under the cortex is implemented on Virtex-5 FPGA experimental platform and the implemented basal ganglia are applied to robot behavior selection.Research results of this dissertation will expand the current studies in behavior-based robotics and contributes to build more intelligent and biologically plausible robots.