A Multi-Agent Advisory Board Reinforcement Learning Framework for Adaptive Cooperative Control
Onur Osman, Tolga Kudret Karaca, Bahar Yalçın Kavuş, Gökalp Tulum, Sajjad Nematzadeh · Algorithms · 2026
This study proposes Advisory Board Reinforcement Learning (AdvB-RL), a cooperative reinforcement-learning framework that integrates multiple advisory neural networks to guide policy optimization. Unlike conventional single-agent architectures, AdvB-RL maintains a set of independently trained advisory networks that contribute to action selection through a dynamic aggregation mechanism. This design preserves diverse experiential knowledge while improving learning stability and the exploration–exploitation balance. The framework is evaluated on three benchmark control tasks, namely LunarLander-v2, CartPole-v1, and MountainCar-v0, using advisory board sizes of 1, 5, and 10 members against a Double Deep Q-Network (DDQN) baseline. The best-performing configuration, 10 AdvB, achieved 270.02 ± 24.74 on LunarLander-v2 versus 227.92 ± 86.02 for DDQN, 497.79 ± 5.18 on CartPole-v1 versus 304.37 ± 144.04, and −103.16 ± 15.46 on MountainCar-v0 versus −130.71 ± 31.64, indicating higher returns together with markedly lower variability. Across the three environments, these results show that increasing the number of advisory members improves both reward consistency and overall robustness, with the 10-member setting providing the strongest performance. Within the tested configurations, the advisory board mechanism remains computationally feasible, while preliminary experiments beyond 10 advisors show diminishing returns relative to added complexity. Overall, AdvB-RL provides a robust and modular alternative to single-policy reinforcement learning for adaptive cooperative control.