Using a Bidirectional Associative Memory and Feature Extraction to model Nonlinear Exploitation Problems
Each day we are faced with a decision of maximizing our resources by using our current knowledge to learn new things. Should we go to the new restaurant that just opened around the corner or stick to an old, reliable favourite? This is known as the exploration-exploitation dilemma and it is at the heart of reinforcement learning. The present study looks at the exploitation half of this problem and aims to implement it in a biologically plausible recurrent associative memory model. In the framework of Artificial Neural Networks, exploitation is observed when the network can iterate through many learned responses and stabilize on the correct one to solve a given task. This is a process akin to being able to switch from a line to a point attractor. More precisely, Bidirectional Associative Memory (BAM) is used to accomplish such tasks where the context dictates which attractor the network should converge to. For simple independent tasks, the BAM is sufficient. However, for overlapping tasks, the task becomes nonlinearly separable. Therefore, the BAM needs an extra unsupervised layer to extract unique features from the inputs. These features combined with input are then sent to BAM where it can learn the different attractors adequately. This network was able to stabilize on the correct responses of tasks that involved time series of varying lengths, overlap, and levels of correlation; the variability one would expect from the real world.