Adaptive Switching Between Self-Selection and Collaborative Selection in Multi-Agent Systems: A Unified Decision-Making Framework

Abstract

Multi-agent systems must repeatedly select actions, solutions, resources, partners, strategies, or organizational configurations while operating under uncertainty and limited computational resources. Existing approaches often emphasize individual decision-making, collaborative decision-making, or predefined coordination policies. However, neither independent selection nor collaborative selection is universally optimal across all task conditions. Independent selection can reduce communication overhead and preserve agent autonomy, whereas collaborative selection can exploit distributed knowledge, complementary capabilities, and collective evaluation. Excessive reliance on either mode may lead to inefficient decisions, unnecessary communication, redundant computation, or insufficient information integration. This paper proposes an Adaptive Self-Selection and Collaborative Selection (ASCCS) framework for multi-agent systems. The central idea is to allow a system to switch dynamically between self-selection and collaborative selection according to task uncertainty, decision confidence, estimated disagreement, expected collaboration benefit, communication cost, computational budget, and historical outcomes. Unlike a fixed policy that always selects independently or collaboratively, the proposed framework treats selection mode as a decision variable that can be optimized during system operation. A hybrid mode is also introduced for cases in which independent candidates are generated first and collaborative selection is activated only when additional evaluation is expected to improve the decision. The framework is formulated as a sequential decision problem and supports rule-based, contextual-bandit, and reinforcement-learning implementations. This paper develops the architecture, mathematical formulation, switching algorithm, evaluation hypotheses, and experimental protocol. Because no new experiments are reported in this conceptual and methodological manuscript, its effectiveness remains to be established empirically.