Autonomy through real-time learning and OpenNARS for Applications

Abstract

This work includes an attempt to enhance the autonomy of intelligent agents via real-time learning.In nature, the ability to learn at runtime gives species which can do so key advantages over others. While most AI systems do not need to have this ability but can be trained before deployment, it allows agents to adapt, at runtime, to changing and generally unknown circumstances, and then to exploit their environment for their own purposes. To reach this goal, in this thesis a pragmatic design (ONA) for a general-purpose reasoner incorporating Non-Axiomatic Reasoning System (NARS) theory is explored. The design and implementation is presented in detail, in addition to the theoretical foundation. Then, experiments related to various system capabilities are carried out and summarized, together with application projects where ONA is utilized: a traffic surveillance application in the Smart City domain to identify traffic anomalies through real-time reasoning and learning, and a system to help first responders by providing driving assistance and presenting of mission-critical information. Also it is shown how reliable real-time learning can help to increase autonomy of intelligent agents beyond the current state-of-the-art. Here, theoretical and practical comparisons with established frameworks and specific techniques such as Q-Learning are made, and it is shown that ONA does also work in non-Markovian environments where Q-Learning cannot be applied. Some of the reasoner's capabilities are also demonstrated on real robotic hardware. The experiments there show combining learning knowledge at runtime with the utilization of only partly complete mission-related background knowledge given by the designer, allowing the agent to perform a complex task from an only minimal mission specification which does not include learnable details. Overall, ONA is suitable for autonomous agents as it combines, in a single technique, the strengths of behavior learning, which is usually captured by Reinforcement Learning, and means-end reasoning (such as Belief-Desire-Intention models with planner) to effectively utilize knowledge expressed by a designer.