Intent-Based Multi-Agent Allocation With Nature-Inspired Online Algorithm

This paper investigates the problem of online agent selection for intent recognition in multi-agent Intent-Based Networking (IBN) systems, where diverse intelligent agents provide intent interpretation under resource and cost constraints. To address dynamic environments, adversarial inputs, and uncertain predictions, we proposeICC-Bandit, an online contextual constrained bandit framework that adaptively selects among heterogeneous agents based on intent input features, cost feedback, and predictive confidence. The term nature-inspired refers to a feedback-reinforced selection mechanism in which agent preference scores are strengthened or weakened by online performance and cost signals, analogous to reinforcement dynamics in swarm coordination, while feasibility is maintained through constraint-aware updates. The proposed method integrates constraint-aware learning with calibrated online regret minimization, ensuring bounded performance loss while respecting system-level constraints such as violation budgets and adaptation latency. We establish theoretical guarantees on dynamic regret and competitive ratio, demonstrating that ICC-Bandit achieves provably efficient learning under adversarial and stochastic conditions. Experiments on three real-world intent datasets show that ICC-Bandit consistently outperforms existing baselines in terms of regret, violation rate, and computational adaptability, achieving up to 44% reduction in dynamic regret and 14% improvement in competitive ratio.