Learning in Dynamic Probabilistic Environments: A Parallel-constraint Satisfaction Network-model Approach