Complex Logical Action-State Prediction

Abstract

This thesis proposes three novel improvements to the Actor-Critic State-Action-Reward-State-Action algorithm while considering potential biologically equivalent mechanisms. The algorithms are optimized via a Particle Swarm Algorithm, tested on a unigram character prediction problem, and evaluated on bit-wise accuracy and character exactness. Some non-unique changes include, kerneling for flexibility in state encoding options, and mixing historical and predictive information into the algorithm's logical input to supplement non-observable elements. The first contribution is a more flexible delta calculation method which better emulates how neurotransmitters are released, recovered, and lost. The second contribution is w.r.t. the implementation of complex weights and states using a trigonometric interpretation, allowing the algorithm to more clearly distinguish between non-observability and non-existence. The last contribution, bounded error, restricts the maximum output magnitude of the logical predictions in a way that improves weight stability and filtration of influence from states with weak relations to the output.