Minimal Requirements for the Emergence of Learned Signalling

Abstract

The emergence of signaling systems has been observed in numerous experimental and realworld contexts, but there is no consensus on which (if any) shared mechanisms underlie such phenomena. A number of explanatory mechanisms have been proposed within several disciplines, all of which have been instantiated as credible working models. However, they are usually framed as being mutually incompatible. Using an exemplar-based framework, we replicate these models in a minimal configuration which allows us to directly compare them. This reveals that the development of optimal signaling is driven by similar mechanisms in each model, which leads us to propose three requirements for the emergence of conventional signaling. These are the creation and transmission of referential information, a systemic bias against ambiguity, and finally some form of information loss. Considering this, we then discuss some implications for theoretical and experimental approaches to the emergence of learned communication.

A Social Approach to Rule Dynamics Using an Agent‐Based Model

“…B), indicating that despite new learners constantly replacing more proficient speakers in the population, there is low interspeaker variation. This may be because turnover combines with memory constraints to create greater overall information loss, which can bolster the emergence of shared conventions (Spike et al., ).…”

“…Agents have temporally constrained memory for each lemma‐inflection combination. Memory limitations are crucial to recovering realistic frequency‐dependent rule dynamics found in natural language (e.g., in corpora; Cuskley et al., ) and have been identified as an essential factor in the emergence of learned signaling in agent‐based models more generally (Spike et al., ). This memory constraint is implemented as a deterministic loss of a lemma‐inflection pairing after a set period of time: If a pairing has not been encountered within a specific time window, d = 100 (regardless of the pairing's weight, w ), the pairing will be forgotten.…”

“…(; see also, Cuskley et al., ) used a model of three‐state dynamics wherein agents have either a regular, irregular, or mixed (optionally regular or irregular) rule for a given lemma. This approach examined both gradual population turnover with biased child learners and memory limitations (also known as, information loss; Spike, Stadler, Kirby, & Smith, ). These factors, combined with a skewed frequency distribution of lemmas, recovered patterns in regularity reminiscent of natural language data (Carrol, Svare, & Salmons, ; Cuskley et al., ; Lieberman, Michel, Jackson, Tang, & Nowak, ), wherein highly frequent forms are more likely to be irregular given certain starting conditions.…”

Communication games: Modeling language evolution through dyadic agent interactions

“…We accomplish this by focusing on two specific implementations that have been applied in language evolution: the Naming Game (NG), which grew out of embodied artificial intelligence research (Steels, 1995), and the Lewis Signaling Game (LSG), which has its roots in game theory (Lewis, 1969). While these are by no means the only possible ways of implementing agent-based communication games (see Spike (2018), ch. 1 for a detailed review), they provide a useful illustration of convergent results, and have been especially influential in language evolution.…”