Computational Modeling of Trust Factors Using Reinforcement Learning

As machine-learning algorithms continue to expand their scope and approach more ambiguous goals, they may be required to make decisions based on data that is often incomplete, imprecise, and uncertain. The capabilities of these models must, in turn, evolve to meet the increasingly complex challenges associated with the deployment and integration of intelligent systems into modern society. Historical variability in the performance of traditional machine-learning models in dynamic environments leads to ambiguity of trust in decisions made by such algorithms. Consequently, the objective of this work is to develop a novel computational model that effectively quantifies the reliability of autonomous decision-making algorithms. The approach relies on the implementation of a neural network based reinforcement learning paradigm known as adaptive critic design to model an adaptive decision making process that is regulated by a quantitative measure of risk associated with each possible decision. Specifically, this work expands on the risk-directed exploration strategies of reinforcement learning to obtain quantitative risk factors for an automated object recognition process in the presence of imprecise data. Accordingly, this work addresses the challenge of automated risk quantification based on the confidence of the decision model and the nature of given data. Additionally, further analysis into risk directed policy development for improved object recognition is presented