Hardware implementation of FAST-based reinforcement learning algorithm

A FAST-based (flexible adaptable-size topology) reinforcement learning chip is implemented in this article. Basically, the FAST is an ART-like (adaptive resonance theory) mechanism. The ART is characterized as one of unsupervised learning neural network models, facilitated to solve stability-plasticity dilemma. The chip is a self organizing architecture that consists of three main structures including similarity, learning, and pruning. Dynamically adjusting the size of sensitivity regions of each neuron and adaptively pruning one of the neurons when an input pattern activates more than one neuron, the chip can preserve hardware resources (available neurons) to accommodate more categories. The clustered result by the implemented chip is then sent to an AHC (adaptive heuristic critic) architecture (emulated by a personal computer) to learn to balance an inverted pendulum system, which is also emulated by the personal computer for verifying the implemented architecture.