Title: Artificial Intelligence Across Disciplines: Research, Innovation, and Intelligent Solutions
Editors: Dr. Subita Bhagat, Dr. A. Balamurugan, Dr. P. Krishna Kumar, and Mrs. S. Nandhini Devi
ISBN: 978-81-69857-83-3
Chapter: 16
DOI: https://doi.org/10.59646/815/16
Author: Kavithamani B
Abstract
Autonomous self-evolving artificial intelligence architectures represent a major operational evolution from static, freeze-trained models toward dynamic, open-world computational systems capable of lifelong adaptation without human re-engineering. This chapter presents an end-to-end framework for self-evolving AI, unifying continuous parameter regularisation, dynamic architectural expansion, generative experience replay, and meta-learned policy optimization. We investigate the implementation of this paradigm within an edge-computing industrial cyber-physical manufacturing testbed characterized by non-stationary data streams, abrupt sensor recalibrations, and evolving mechanical failure signatures. By incorporating a Riemannian manifold regularizer alongside neural architecture search controllers, the proposed system mitigates catastrophic forgetting while autonomously allocating new representational capacity when statistical concept drift is detected. Evaluated over 360 operational days across 50 heterogeneous sensor nodes, the architecture sustained an aggregate operational classification accuracy of 98.4%, prevented historical task degradation to within 1.2%, and cut retraining compute overhead by 62.8% relative to full periodic retraining routines.
Keywords: Self-Evolving AI, Continual Learning, Catastrophic Forgetting, Dynamic Architecture Expansion, Generative Replay, Meta-Learning.