A probabilistic RAM controller with local reinforcement learning

A probabilistic RAM controller comprising 128 local learning probabilistic RAMs is proposed. The probabilistic RAM (pRAM) is a highly nonlinear and stochastic device possessing neuron-like properties. A reinforcement learning rule with a local learning environment has been implemented. Two non-learning pRAMs have been used to provide the environmental information for the reinforcement learning rule. Therefore, the learning curve of the pRAMs can be controlled individually. The pRAM controller uses an external RAM chip for pRAM memory and allows 128 pRAMs to be controlled by using one custom VLSI device. The pRAM controller has four transputer-like data links, which enable interfacing of the other pRAM controllers. Hence, the size of the neural networks can be scaled to accommodate different applications