Article Details
  • Published Online:
    May  2025
  • Product Name:
    The IUP Journal of Telecommunications
  • Product Type:
    Article
  • Product Code:
    IJCT040525
  • DOI:
    10.71329/IUPJTC/2025.17.2.63-82
  • Author Name:
    Kaushik Bar
  • Availability:
    YES
  • Subject/Domain:
    Engineering
  • Download Format:
    PDF
  • Pages:
    63-82
Volume 17, Issue 2, May 2025
Self-Improving Agentic AI Through Bayesian Meta-Learning
Abstract

The paper proposes a novel framework for developing self-improving agentic artificial intelligence (AI) using Bayesian meta-learning (BML). Traditional agentic AI systems often struggle with adapting to dynamic and uncertain environments, limiting their ability to autonomously refine their decision-making strategies. By integrating BML, the proposed model enables agents to perform probabilistic reasoning, quantify uncertainty, and adapt their policies based on experience. The framework employs a hierarchical Bayesian approach, where priors are iteratively updated using real-world feedback, leading to continuous selfimprovement. Experiments were conducted in simulated environments, including robotics control tasks and autonomous driving scenarios. The results demonstrate that the proposed model achieves faster adaptation, higher decision quality, and reduced uncertainty, compared to conventional reinforcement learning (RL) and meta-learning baselines. Additionally, the model’s ability to generalize across diverse tasks was evaluated using metrics such as adaptation speed, cumulative reward, and uncertainty reduction. The paper contributes to the advancement of autonomous AI systems, paving the way for more reliable and selfsufficient intelligent agents in real-world applications.

Introduction

The rise of agentic artificial intelligence (AI) systems—capable of autonomous decision making, adaptation, and self-improvement—marks a significant shift in the development of intelligent systems. Traditional AI models rely heavily on predefined rules or extensive supervised learning, limiting their ability to generalize across tasks and environments. In contrast, agentic AI aims to function with minimal human intervention, continually refining its strategies based on real-time feedback. However, a key challenge in deploying such systems is their ability to manage uncertainty while making high-stakes decisions. This paper explores the integration of Bayesian Meta- Learning (BML) as a means to equip agentic AI with self-improving capabilities, allowing for adaptive decision making and continuous learning.