MMAO-Cls optimizer achieves compact features with near-top accuracy

In a new arXiv preprint, Jinliang Xu and Liping Ma introduce MMAO-Cls, a metabolic multi-agent optimization framework designed for joint feature selection and classifier tuning. The approach extends the Metabolic Multi-Agent Optimizer (MMAO) into a mixed-space search where each agent encodes both a binary feature mask and classifier hyperparameters. Key innovations include mapping private energy, communal budget, role drift, and lifecycle turnover to the accuracy-complexity tradeoff of wrapper learning, plus feature-budget adaptation derived from feature-information priors and regularization against overfitting. The method was evaluated on seven standard tabular datasets with three seeds each, comparing against RandomSearch, GA-lite, PSO-lite, and a no-sharing ablation.

Results show MMAO-Cls achieving a mean validation objective of 0.9433 (second best behind GA-lite at 0.9446) and a held-out test score of 0.8882, improving over RandomSearch (0.8808) and GA-lite (0.8857), while remaining close to PSO-lite (0.8874) and the no-sharing ablation (0.8900). Most notably, MMAO-Cls produced the most compact selected feature subset with a mean feature ratio of 0.4881, meaning it used about half the features of other methods while maintaining competitive accuracy. However, pairwise statistical tests showed these margins are not yet significant, leading the authors to a conservative conclusion: MMAO-Cls clearly supports classification applicability and compact mixed-space search, but the specific advantage of communal sharing is not decisively isolated. The work suggests practical benefits for building lean, interpretable classifiers in resource-constrained settings.