Abstract
While computation power has increased and the statistical machine learning methods have made substantial advancement, many problems that would benefit from real-time interpretation have not exploited their combined strengths. For instance, the problem of gathering data from the environment and transforming it into knowledge as well as updating the knowledge as new data become available. Currently, with substantial expressivity and moderate computational cost, high-level languages or first-order predicate logic or model-based machine learning are used for static representation of knowledge, that is used for reasoning and inferring. In this dissertation, we address how an entity dynamically gather knowledge from environmental data and use that for inferring evolving events and dynamically update the current knowledge. We develop theoretical and empirical solutions using Description Logic representation and reasoning, and General Value Functions in Reinforcement Learning. The proposed solutions dynamically extract low-level knowledge from available data and update the high-level knowledge, which is used to predict the evolving future events. We show its applications in three real world domains: 1) RoboCup 3D Soccer Simulation environment, 2) High-throughput screening, and 3) Axon regeneration.