Decentralised Emergence of Robust and Adaptive Linguistic Conventions in Populations of Autonomous Agents Grounded in Continuous Worlds

TL;DR

The paper tackles how a population of autonomous agents can autonomously bootstrap a human-like linguistic convention that maps word forms to grounded, continuous concept representations. It adopts a decentralised language-game framework where each agent builds an inventory of words tied to channel-wise Gaussian concept representations, and where word meaning evolves through local interactions, discriminative production, and entrenchment-based alignment. Key contributions include grounding concepts in a multi-sensor continuous space, achieving high communicative success and coherence across diverse domains (CLEVR, WINE, CREDIT), demonstrating compositional generalisation with CLEVR CoGenT, and proving robustness to heteromorphic sensors, sensor defects, perception differences, and continual learning without catastrophic forgetting. The results suggest practical applicability for robust, adaptive emergent communication in real-world multi-agent systems, with implications for scalable human-agent and agent-agent communication in dynamic environments. The approach relies on concrete metrics, including the channel-weighted similarity $sim(c,X_S)$ and Hellinger-based concept similarity, and leverages online updates like Welford's algorithm to maintain statistically grounded representations.

Abstract

This paper introduces a methodology through which a population of autonomous agents can establish a linguistic convention that enables them to refer to arbitrary entities that they observe in their environment. The linguistic convention emerges in a decentralised manner through local communicative interactions between pairs of agents drawn from the population. The convention consists of symbolic labels (word forms) associated to concept representations (word meanings) that are grounded in a continuous feature space. The concept representations of each agent are individually constructed yet compatible on a communicative level. Through a range of experiments, we show (i) that the methodology enables a population to converge on a communicatively effective, coherent and human-interpretable linguistic convention, (ii) that it is naturally robust against sensor defects in individual agents, (iii) that it can effectively deal with noisy observations, uncalibrated sensors and heteromorphic populations, (iv) that the method is adequate for continual learning, and (v) that the convention self-adapts to changes in the environment and communicative needs of the agents.

Figures (4)

• Figure 1: Examples of emerged concepts for the CLEVR (a), WINE (b) and CREDIT (c) datasets.
• Figure 2: Evolutionary dynamics during the training phase of the CLEVR experiment: degree of communicative success, degree of linguistic coherence and average linguistic inventory size as a function of the number games that are played.
• Figure 3: Evolutionary dynamics during the training phase of the CLEVR experiment in which each agent loses access to 1 or 10 sensors after 500,000 games.
• Figure 4: Evolutionary dynamics during the continual learning experiment, in which a model is first trained on CLEVR and then on WINE.