A Working Theory of a Learned Model in a Partially Observable Environment for Cognitive Decision-Making

Abstract

To survive in our unpredictable, evolving world, cognitive beings learn to make decisions with the limited knowledge of the world they process. Reflective of an individual's view of the world, a cognitive decision-making model is explored in a partially observable, stochastic environment. The cognitive model uses the Partially Observable Markov Decision Process problem formulation, which is a framework for neurological models and considered implementable in neural circuitry [26] [16]. To structure a planning model comparable to that of DeepMind's MuZero in a partially observable environment, a belief function will translate the observations to a vector of belief states that will be discretized so as to be used as the observations of a MuZero-based machine learning algorithm [29]. The belief states are computed recursively from the previous belief state using Bayesian inference. Bayes rule is thought to capture the neurological and cognitive levels of reasoning [26]. Components of the planning, training, and action methods of the cognitive model will follow those of MuZero. The model could then be trained and act, in way parallel to that of MuZero, in a partially observable environment. Cognitive insights from a model structured in this form and additional considerations are discussed.