Hybrid Computational Models for Skill Acquisition

Abstract : The goal of this research was to develop a hybrid real time problem solving architecture that couples symbolic planning methods with connectionist reinforcement learning methods. The advantage of this hybrid architecture is that it can immediately achieve reasonable performance, because the symbolic planning system can quickly develop an acceptable control policy, but it can also gradually achieve optimal real time performance, because the reinforcement learning system will eventually converge on a near optimal policy. Many DoD problems would benefit from the ability to perform near optimal real time control of complex systems.