Evolution of reflexive signals using a realistic vocal tract model

Abstract

We introduce a model of the evolution of reflexive primate signals that incorporates a realistic vocal tract model for generating the signals. Signaler neural networks receive signal types as inputs and produce vocal tract muscle activations as outputs. These muscle activations are input to a model of the primate vocal tract, generating real sounds. Receiver neural networks receive spectrograms of these sounds as inputs and produce signal type classifications as outputs. Incorporating a realistic vocal tract has a substantial effect on the types of signals that can evolve. Compared to a model with abstract signals, the realistic model signals are more similar and have more correlated elements. The realistic, embodied model also exhibits more variability in rate of adaptation, usually adapting more slowly. This may be explained by the more jagged fitness landscapes in the realistic model. The realistic signals also tend to be quiet. Environmental noise results in louder signals but makes the evolutionary process even slower and less robust. These results indicate that signal evolution with a more realistic genotype–phenotype mapping can differ substantially from evolution with abstract signals. Including realistic signal generation mechanisms may enable computational models to provide greater insights into natural signal evolution.