Brian Skyrms’s latest book, Signals: Evolution, Learning, and Information, begins with a very simple game. Two individuals need to coordinate an action with some state of the world. One individual, the sender, learns the state of the world and can choose some signal to send. The other, the receiver, perceives the signal, but not the state of the world, and must choose an action. They each get paid if and only if the action is appropriate for the state of the world. The signals have no natural salience or intrinsic “meaning,” yet coordinated use by sender and receiver can result in a successful solution. How can such solutions arise spontaneously via natural processes?
This sender-receiver game was originally presented by David Lewis in Convention (Cambridge, MA: Harvard University Press, 1969) to explore the nature and conventionality of linguistic meaning. But Lewis does not provide a satisfactory answer to the question of how signaling can evolve. This is the central focus of Skyrms’s book. He uses the contemporary tools of the social and biological sciences to investigate how the natural processes of evolution or learning can generate signaling systems in the sender-receiver game. Skyrms also explores the significance of this investigation for our understanding of linguistic meaning, deception, logic, information processing, and social interaction. Along the way, he presents concise, clear, and effective introductory chapters to evolutionary game theory, learning dynamics, and information theory. The result is an insightful book that prompts a change in the way the old issues are viewed and raises a host of important questions and new avenues for investigation.
Skyrms begins by introducing the basic signaling game framework before surveying a number of fascinating signaling systems in nature ranging from chemical signals in bacteria to the syntax of chickadee songs. The first set of results presented suggests that evolutionary processes can relatively easily achieve signaling systems in the sender-receiver game. However, it quickly becomes apparent that the sender-receiver game is deceptively simple with complexities and nuanced issues lurking just under the surface. Beyond the simplest settings (two states, two signals, two actions, with each state occurring exactly half of the time), evolution does not always deliver signaling systems and can often result in pooling or partial-pooling equilibria that are less than optimal. Skyrms considers a number of variations on this simple game, some involving too many signals, or too few, conflicting interests, signaling in networks, and so on. The evolutionary results include the use of signals that resemble phenomena such as synonyms, categories, deception, and even logical inference.
Some may object that the signals in the simple sender-receiver game are not sufficiently similar to sentences or propositions, and so analyzing these signals cannot help us understand robust semantic notions such as meaning. In these signaling games, there is just behavioral use of the terms (by entities as simple as bacteria in some cases). Thus, the models Skyrms is considering clearly lack the important intentionality required if they are to help us understand the nature of meaning and semantic content.
Skyrms largely sidesteps these issues by focusing on the more general notion of information. Drawing on several ideas from information theory, he provides a simple measure of the quantity of information a signal carries about a state (involving the degree to which the signal increases the probability of the state). This measure of information can be used to develop a representation of the informational content of a signal. Here the standard understandings of content (i.e., propositional content) become special cases of informational content.
One interesting application of this framework is to the notion of deception. Deception, for Skyrms, is just misinformation that benefits the sender at the detriment of the receiver. And misinformation is a signal that raises the probability of an untrue state. This means, we can have deception without any “intent” to deceive or in organisms without much cognitive sophistication (e.g., predatory fireflies). There are also examples of half-truths that carry both information and misinformation, but may not count as deception if the receiver benefits. In some signaling games, the optimal evolutionary outcomes involve a kind of universal deception where all signals are deceptive. In the models Skyrms presents on deception there is no obvious intentionality. Thus, one might view Skyrms as a kind of reductionist or deflationist with respect to deception (or more general notions of “meaning”). However, I think a more accurate characterization of Skyrms’s view is simply that information, meaning, deception, and so forth do not require mental states, cognition, or some robust form of intentionality. Some special instances of signaling may rest heavily on those kinds of cognitive aspects. But, these are just part of a more general set of phenomena, which is the real target of Signals.
In addition to exploring signaling games in a traditional evolutionary setting, Skyrms also examines various models that involve agents learning to signal. He presents results (some analytic, some from computer simulation) demonstrating that a variety of simple learning rules can generate efficient signaling spontaneously. Some rules are better than others at avoiding the pitfalls of partial-pooling equilibria where signaling is less than optimal. Sometimes even remarkably simple learners are exceptionally good at learning to signal. For instance, simple stimulus-response reinforcement learners can very often reach efficient signaling systems, though they do have certain difficulties with signaling about rare events or handling a large number of states, signals, and actions in a maximally efficient manner. A very simple best-response learning rule that best responds to the last play of the partner with some probability can achieve optimal signaling systems 100% of the time, even in the harder cases.
The chapters on learning to signal present a plethora of results for a wide variety of learning rules and generalizations of the sender-receiver game. But, there are many points where Skyrms does not present a detailed overarching view and leaves issues open or questions unanswered. For example, Skyrms does not tell us what happens if two different kinds of learners are interacting. Imagine a sender who is using the best-response rule and a receiver that learns via reinforcement: does the gradual probabilistic shifting of reinforcement learning interfere with the rapid shifting of a best responder? Though Skyrms does not address this question, it is possible to reconstruct Skyrms’s models and set one of the two players as a reinforcement learner and the other as a best-response learner. Prompted by curiosity, I ran computer simulations of this scenario myself. I used a two-state, two-signal, two-act game where one state occurred 90% of the time—a scenario where reinforcement learners have some difficulty reaching optimal signaling systems. It turns out that one best responder and one reinforcement learner are better at achieving optimal signaling than two reinforcement learners, and are nearly as good as two best responders.
There are also other points at which Skyrms is content to present a partial view with suggestive results. For instance, in chapter 11, he applies the signaling game framework to the notion of logical inference, which he views as a subset of information processing. However, he only shows how to understand “a few rudiments of this process,” and the chapter ends on a suggestive note. While some may find this unsatisfying, I think this is a virtue of the book. These open-ended discussions and incomplete answers one finds in Signals are simply a result of the vast number of possibilities the signaling-game framework allows us to investigate. Skyrms illuminates these possibilities throughout the book, raising a number of new and interesting questions. Often, he only provides partial answers, but he also provides the tools necessary for a detailed and precise investigation.
One of the most innovative ideas in Signals is the idea that the game itself may evolve or change. This idea is relevant to a wide range of possible applications of game theory, but it is particularly important for the evolution of communication. After all, a full account of the evolution of language should include not just establishing signaling systems but also the origin (or invention) of the signals. This requires modifying the original game so that the players must first invent signals and then learn how to use them. To address this, Skyrms proposes a model that is able to represent an evolving signaling game as the agents (represented as a variation of simple reinforcement learners) invent new signals during the course of interaction. Introducing this possibility of invention allows the agents to avoid the evolutionary traps of suboptimal signaling. More impressively, Skyrms finds that individuals in this model can begin with no signals and invent a signaling system from scratch that almost always enables near perfect communication.
The book ends with a set of chapters exploring the connections between signaling and networks of individuals. There are a number of interesting models examined including signal-mediated teamwork, dialogue, and dynamic networks of signaling. The results here are primarily exploratory but are very intriguing.
In summary, Skyrms succeeds in showing that effective signaling systems can arise from very simple processes such as basic reinforcement learning or natural selection. But in doing so, he opens new avenues for research and raises many interesting questions. I think one of the most important contributions of Signals is in demonstrating the capabilities of a new research program. Overall, Skyrms’s Signals is a thought-provoking introduction to an area of research that is rapidly developing and promises to reshape the way we think about a number of issues in philosophy of language and epistemology.
