Getting Mixed Signals: Exploring the Evolution of Disjunctive Signaling Games

This post is written by Peter Fetros, an undergraduate computer science research assistant at UI working with James Foster and Bert Baumgaertner

Peter Fetros explaining a section of their research poster presented at an undergraduate research conference

Peter Fetros explaining a section of their research poster presented at an undergraduate research conference

Signals are all around us. Most organisms use signals in order to communicate with one another. They might use them to tell others where the best food is, or possibly to warn them about approaching predators. However not very much is known about the evolution of these signals. How and why some signals are used while others are not, and how organisms decided to use these signals. I’ve always been interested in the ways that things communicate, from tiny organisms who emit chemicals to ward of predators, to humans and the thousands of different ways we choose to communicate information to one another. However, as a Computer Science Student I’m even more interested in finding a way to model these interactions in a meaningful way that will give us insight into how meaningful signals can evolve. Our current research explores how signals evolve over time and what that evolution looks like. We do this by modeling the interaction of “players” in something called a Lewis Signaling-Game.

A Lewis Signaling-Game in its most simple form consists of 2 players, a Signaler and a Receiver. In this game the player, or “agent”, who is designated as the Signaler first observes some world-state. The Signaler then chooses a signal to send to the Receiver based on the world-state it observed. After the Receiver receives the signal from the Signaler, The Receiver then chooses to perform an action based on that signal. If this action is the correct action for the world-state that the Signaler originally observed, they both get a payoff.

An example of this process in nature can be compared to the calls of Vervet monkeys. These monkeys have different vocal calls (signals) for when they see different types of predators, for example leopard, eagles, or snakes. When a Vervet monkey sees a predator (an observation) it will choose the call appropriate for this predator (a signal). It will make the call and then another monkey who cannot see the predator will hear it (Receiver). The monkey who hears the call will then have to decide how to properly avoid this unseen predator (choosing an action). In this example that might be climbing a tree in order to get away from a leopard prowling on the ground, running away from the comparably slow snake, or even hiding in a bush so the eagle can’t reach it. If the Vervet monkey chooses the correct action for the type of predator, it will get a payoff (not being eaten), However if it chooses the wrong action for the type of predator it will not get a payoff (get eaten).

To model these interactions on a large scale we model the Signaling-Game in a programming language called NetLogo. In our models we have a large population of agents that all pick a partner and then play the Signaling-Game. They start out not knowing what signal is correct to use for what world state it observed, as well as what action they should perform when they receive a signal. However, when they do choose the right action based on the world state the observer sees, they get a payoff in the form of an increase in their preference for choosing that signal again (if the player was a Signaler), or choosing that action (if the player was a Receiver) for that specific observation. After each game the agents randomly pick new partners and then play again. Eventually they do this enough so that they all agree on the same meaning for signals. When this happens we call it a signaling System.

There has already been some research into the basic Lewis Signaling Game and Signaling Systems. So what we are currently exploring is how these Signaling Systems evolve when we have observations that are disjunctive. Disjunctive observations are observations that the Signaler can sometimes make in which it doesn’t know the true state of the world. It might be World “A” or World “B”.

An example of this using the Vervet Monkeys might be when a monkey sees a rustling bush and doesn’t know whether it’s a leopard or a snake. Because it doesn’t know which it really is, it must somehow let the other monkeys know that there is a ground predator nearby but it isn’t sure what kind, all it knows is it isn’t an eagle. There are several ways it might do this. It may send a new type of call that means “Leopard or Snake” or it may just guess itself, and make the call for leopard or the call for snake depending on what it decides might be in the bush.

Ternary plot detailing the population preferences for the agents playing as Receivers

Ternary plot detailing the population preferences for the agents playing as Receivers

So far in our simulation we have discovered several different paths the evolution of these Signaling Systems may take. Sometimes the Signalers use a new kind of signal and sometimes they use the old signals while just making a guess at what the world state might be. We plot the results from the simulations on ternary-plots. These plots show the population’s average preference for a particular signal or action based on either the world-state it observed or the signal it received respectively. These graphs allow us to see how a particular preference changed over time at the population level.

If you have any feedback or questions about this research, please contact

Peter Fetros (, James Foster (, Bert Baumgaertner ( or Kelly Christensen (

This entry was posted in BEACON Researchers at Work and tagged , , , . Bookmark the permalink.

Comments are closed.