Modular SRV reinforcement learning: an architecture for nonlinear control

Demonstrates the application of a hybrid reinforcement-modular neural network architecture to nonlinear control problems. Specifically, the method of action-critic reinforcement learning, modular neural networks, and winner-takes-all updating are combined. This provides an architecture able to both support temporal difference learning, and probabilistic partitioning of the input space. Furthermore, the number of partitions required a priori of the input space is far lower than that previously in the norm. Application of this methodology to the pole-balancing problem demonstrates superior partitioning of the input space, bettering that of equivalent BP networks; whilst avoiding the learning to learn nothing effect, as is often the case with temporally dependent problems.