The Non-Axiomatic Reasoning System (NARS) is a logic-based AI architecture that supports explainable reasoning, explicit knowledge representation, general-purpose processing, and online learning. Despite these advantages, scaling NARS up to complex application domains remains difficult because strong autonomous performance typically depends on extensive manual design. Often in practice, one or more human designers must study the target domain, craft background knowledge, tune numeric system parameters, and sometimes modify the system architecture to obtain a high-performing system configuration. While pre-training is standard in neural approaches, no general method exists to automatically pre-optimize NARS systems for learning and performing in sensorimotor domains. This dissertation introduces a methodology and genetic encoding to optimize NARS configurations using evolutionary algorithms. In contrast to neural models optimized via gradients, NARS configuration spans symbolic knowledge structures, discrete mechanism choices, and continuous parameters, yielding objectives that are not readily differentiable; this motivates the use of evolutionary search. The proposed evolutionary approach serves a role analogous to pre-training in neural systems: it produces a task-adapted initialization of background knowledge and system configuration prior to deployment to achieve better performance and learning abilities. Given an agent's sensorimotor vocabulary (i.e., the sets of its sensors and motors), the encoding generates and mutates background knowledge (``instincts'') intended to bootstrap useful behavior. In addition, the encoding optimizes numeric meta-parameters that influence NARS architecture and learning dynamics. To increase expressive power, the encoding was extended with support for the implantation and learning of variables and compound structures, improving performance in scenarios that benefit from more expressive knowledge patterns. The proposed method was evaluated across four domains spanning perception, control, multi-agent interaction, and planning: (1) a multi-agent wheeled-robot ecosystem environment, (2) a 2D multi-agent grid-world survival environment, (3) a BlocksWorld planning puzzle scenario, and (4) 3D soft voxel robot locomotion (including both pure NARS and a neuro-symbolic variant). Experiments were designed to assess the effects of evolved instincts alone, runtime (``lifelong''/``online'') learning alone, and the combination of both, establishing the contribution of each method to task performance. Across these domains, the evolutionary process consistently discovered NARS agents that achieved substantial improvements in task performance. Agents with configurations evolved using the basic encoding outperformed fixed-configuration baselines, and the extended encodings with variables and compounds provided additional gains on select tasks. Because NARS maintains explicit internal knowledge structures, evolved behavior was able to be further analyzed through inspection of the system's memory contents to connect agent actions to the learned or implanted reasons that produced them. Overall, this dissertation presents the first genetic encoding for NARS that can jointly evolve meta-parameters and structured symbolic knowledge for sensorimotor domains. It empirically shows that evolutionary optimization with the proposed genetic encoding yields substantial performance gains over fixed-configuration baselines, and that extensions to the encoding can improve performance over the basic encoding in certain scenarios benefiting from generalization and compounds. Evolved agents can be deployed in application settings and also serve as analyzable artifacts for research and development, providing a path toward more robust general-purpose logical reasoning systems.
Christian Geoffrey Hahm (Thu,) studied this question.