Loading...
Non-Axiomatic Reasoning and Causal Explorative System (NARCES), a Hybrid Symbolic Reasoning with Sub-symbolic Learning for Autonomous Intelligent Agents
Isaev, Peter
Isaev, Peter
Citations
Altmetric:
Genre
Thesis/Dissertation
Date
2025-08
Advisor
Committee member
Group
Department
Computer and Information Science
Permanent link to this record
Collections
Research Projects
Organizational Units
Journal Issue
DOI
https://doi.org/10.34944/e9tk-3e53
Abstract
In Artificial Intelligence, reasoning can be distinguished into symbolic and sub-symbolic. Symbolic reasoning makes use of formal language for semantic knowledge representations, as well as utilizes logic for inference, and system control for processing input and derived knowledge. The paradigm examples of sub-symbolic reasoning are biological and artificial neural networks. Given the signals from sensory neurons, our nervous system identifies whether a particular flying object is a dangerous wasp or just a harmless fly. However, nowhere during the process `symbols' do occur, instead, there are only neurons, synapses, electrical signals, etc. Symbols such as `black’, ‘wasp’, or `dangerous’ emerge as high-level descriptions of the process at best.
Recently the focus has been shifted toward combining the two approaches to provide hybrid solutions for practical and effective real-time reasoning, thereby we present NARCES (Non-Axiomatic Reasoning and Causal Explorative System), a novel integrated multi-component system for autonomous intelligent agents, which consists of two main parts: MeTTa-NARS and NACE. MeTTa-NARS is the advanced stand-alone symbolic reasoning system implemented in MeTTa functional language that follows NARS model of intelligence. MeTTa-NARS employs a concept-centered semantic knowledge representation, Non-Axiomatic Logic (NAL) with Narsese formal language and sophisticated control, all of which make it capable of executing various forms of reasoning including deductive, inductive, and abductive reasoning. It also handles uncertainty, revises its knowledge based on new evidence, and learns cause-effect relations from observation in real-time, rather than solely using a predefined set of rules. The second component, Non-Axiomatic Causal Explorer (NACE), is an example of experiential learning, which is a novel sub-symbolic system that showcases high data efficiency in learning and sequential decision-making tasks in grid world environments. NACE is developed with uncertainty in mind and thereby is driven by curiosity. It is capable of capturing logical causal relation induction from observations in real-time using some aspects of NAL as well as causal relation of transitions between parts of states. It iteratively learns an emerging behavior creating structures to allow more robust and efficient decision-making and planning and does not feature encodings of high-level semantic knowledge. Our comprehensive experiments show that NACE performs considerably more data efficient while reaching respectably high mean reward than most state-of-the-art Reinforcement Learning models in Minigrid environments.
Bringing two of the components into NARCES prototype, we chose to work toward a unified foundation of symbolic and sub-symbolic reasoning. It is clear, that we need semantics that add meaning to NACE’s actions and cell values and at the same time MeTTa-NARS will benefit from efficient and robust sub-symbolic processing. Thereby with this framework in mind, we try to resolve a fundamentally important question: how can we obtain a combined system where semantics as in NARS and its high-level reasoning can be effectively interfaced with an array-based world representation for effective sensorimotor processing? This entails the use of a 2D array with values such as those obtained from a grid world, while the NAL representations will allow for human-provided background and derived knowledge to be incorporated and high-level semantic objectives to be pursued.
Furthermore, we showcase the Explorer, a robotic agent that leverages NARCES and operates within the complete ROS2 framework, incorporating simulated environments provided by the Gazebo robotics simulator. By employing industry-standard simulation tools and the ROS2 framework, the Explorer emerges as a promising system that is prepared for testing on actual robots and for deployment in real-world settings.
Description
Citation
Citation to related work
Has part
ADA compliance
For Americans with Disabilities Act (ADA) accommodation, including help with reading this content, please contact scholarshare@temple.edu