Heuristic dynamic programming with internal goal representation

In this paper, we analyze an internal goal structure based on heuristic dynamic programming, named GrHDP, to tackle the 2-D maze navigation problem. Classical reinforcement learning approaches have been introduced to solve this problem in literature, yet no intermediate reward has been assigned before reaching the final goal. In this paper, we integrated one additional network, namely goal network, into the traditional heuristic dynamic programming (HDP) design to provide the internal reward/goal representation. The architecture of our proposed approach is presented, followed by the simulation of 2-D maze navigation (10*10) problem. For fair comparison, we conduct the same simulation environment settings for the traditional HDP approach. Simulation results show that our proposed GrHDP can obtain faster convergent speed with respect to the sum of square error, and also achieve lower error eventually.