Conversational AutoML with Multi-Agent Architecture and Monte Carlo Tree Search for Hyperparameter Optimization

Automated Machine Learning (AutoML) reduces the technical requirements for applying machine learning, yet existing systems still demand technical expertise, optimize inefficiently, and rarely support conversational use. We present a conversational AutoML system with a five-agent architecture coordinated via natural language, tuning Random Forest and gradient-boosting models (XGBoost, LightGBM, and CatBoost) with Monte Carlo Tree Search (MCTS) and Bayesian optimization under repeated cross-validation conditions. Its distinctive component is an ensemble coordinator: specialist optimizer agents tune model families concurrently, and a coordinator fuses their outputs with the strongest baseline into a stacking ensemble, deploying it only when it improves on the best single model by cross-validation. This separates two measurable benefits: a speed-up from parallel search and an accuracy gain from ensemble fusion. On six benchmark datasets (five classification and one regression), the multi-agent coordinator is best or tied on every dataset and the only configuration with a positive average gain over defaults (+0.20 pp, versus −0.69 pp for single-agent MCTS); its search is 2.3× faster than the single-agent Bayesian baseline in aggregate, and it is systematically better than single-agent MCTS across datasets (Wilcoxon signed-rank, p = 0.031). While individual dataset gains are observed, none reach statistical significance when adjusted for multiple comparisons. For classification tasks, we additionally report F1-score, precision, recall, and AUC-ROC (Area Under the Receiver Operating Characteristic Curve).