Mixture of personality improved spiking actor network for efficient multi-agent cooperation.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Computational Neuroscience

Background:

• Multi-agent reinforcement learning (MARL) faces challenges in adapting to new, unseen partners, a problem known as poor new-player-generalization.
• Existing deep-learning algorithms in MARL often neglect the role of theory-of-mind (ToM), hindering adaptive cooperation.
• Human cognition utilizes intuitive personality prediction to navigate social interactions, offering a model for AI.

Purpose of the Study:

• To develop a biologically-plausible MARL algorithm that improves cooperation with unseen partners.
• To integrate theory-of-mind principles into a spiking neural network architecture for enhanced generalization.
• To address the poor new-player-generalization problem in MARL.

Main Methods:

• Proposed a novel algorithm named Mixture of Personality (MoP) improved Spiking Actor Network (SAN).
• The MoP module utilizes a determinantal point process for simulating personality formation and integration.
• The SAN module incorporates spiking neurons for efficient reinforcement learning.

Main Results:

• The MoP-SAN algorithm achieved higher performance in cooperative tasks, both with and without unseen partners.
• Ablation studies confirmed the significant contribution of the MoP module to the SAN's learning capabilities.
• The algorithm demonstrated superior performance compared to conventional deep actor networks.

Conclusions:

• The proposed MoP-SAN algorithm effectively enhances adaptive multi-agent cooperation, particularly in scenarios involving unseen partners.
• Integrating theory-of-mind inspired personality modeling into spiking neural networks offers a promising direction for MARL research.
• The findings suggest that simulating intuitive personality prediction is crucial for robust generalization in MARL systems.