Interpretable Emergent Language Using Inter-Agent Transformers

TL;DR

This work investigates how transformer-based communication can yield human-interpretable emergent language in cooperative MARL. By separating speaker and listener roles and training with decentralized PPO, DIAT learns discrete vocabularies and meaningful embeddings that encode observations into symbolic protocols. Across simple symbolic, shapes/colors, and spatial tasks, the approach demonstrates interpretable language emergence and interpretable embeddings, offering transparency in complex multi-agent interactions. The results suggest that self-attention-based inter-agent communication can produce scalable, human-understandable protocols with potential applications in more complex relational environments.

Abstract

This paper explores the emergence of language in multi-agent reinforcement learning (MARL) using transformers. Existing methods such as RIAL, DIAL, and CommNet enable agent communication but lack interpretability. We propose Differentiable Inter-Agent Transformers (DIAT), which leverage self-attention to learn symbolic, human-understandable communication protocols. Through experiments, DIAT demonstrates the ability to encode observations into interpretable vocabularies and meaningful embeddings, effectively solving cooperative tasks. These results highlight the potential of DIAT for interpretable communication in complex multi-agent environments.

Figures (7)

• Figure 1: Overall Architecture of Communication Scheme
• Figure 2: Tables and accuracy graphs for Trials 1--3
• Figure 3: Shapes in Observation Space.
• Figure 4: Accuracy curve during training for the shapes-and-colors experiment.
• Figure 5: t-SNE of the learned embeddings for the speaker and listener.