1. LEWIS AND BEYOND

Brian Skyrms's Signals has the virtues familiar from his Evolution of the Social Contract (Reference Skyrms1996) and The Stag Hunt (Reference Skyrms2003). He begins with a very simple model of agents in interaction, and in a series of brief and beautifully clear chapters, this model and its successors are explored, elaborated, connected and illustrated through biological theory and the social sciences. Signals borrows its core model from David Lewis: it is Lewis's signalling game. In this game, two agents interact. One agent can observe which of two equi-probable states the world is in, but that agent cannot act directly and profitably on that information. However, the informed agent can act in a way that will be perceptually salient to a second agent: say, by raising a red or a green flag. The second agent does have the capacity to respond appropriately to each state of the world. If that second agent chooses the right option, given the state of the world, both are rewarded. If the second agent fails to choose the right action, neither are. Obviously, the two agents are best off if they have a practice in which the informed agent regularly chooses a distinct, salient cue in response to each of the two world states, and in which the powerful agent uses that cue to select the rewarding act. Less obviously, agents with simple trial and error learning capacities can learn to signal and respond: neither explicit negotiation nor cognitive sophistication are required. Likewise, if individual agents do not have the capacity to learn, but if they breed true but with some variation, the evolutionary version of trial and error learning can take a population to one of the signalling system equilibria.

The lesson of the base case is that communication does not pose a chicken and egg problem. Signalling systems can emerge and be stable between agents who have no prior communicative capacities and only fairly modest capacities to learn. Indeed, the multi-generational route to signalling does not depend on any individual capacity to learn. Nature did not have to wait for intelligence before building communication. Coordination through signals between agents is a resource from which social or biological complexity can be built; it does not itself depend on complexity. To use Dennett's vivid term, signalling is a crane, not a skyhook (Dennett Reference Dennett1995).

Much of the book elaborates this basic conclusion. Signalling can still emerge, and be stable, once we relax some of the extreme idealizations of the basic model. For example: it still emerges in the face of some level of error; in the face of a demand to source information from more than one agent; despite informational bottlenecks (not enough potential signals for each state that matters). Signalling systems can establish even when agents do not have identical interests. Skyrms also explores the effects of different learning strategies. Trial and error learning comes in different flavours, and while they all suffice for the emergence of signals in the simplest case, they are not equally efficient in more elaborate situations.

Some variations from the basic case do block the emergence of signalling. Instead we get what Skyrms calls ‘complete or partial pooling equilibria’ in which signals are ignored, or never sent (or both). Perhaps the most intuitively obvious case is when we dramatically relax the assumption that states of the world are equiprobable. If (say) state 2 is much more likely than state 1, it can pay the responder to always act as if it is 2, rather than pay the trial and error costs of learning which of the informed agent's signals means ‘2’. But even when we factor in complications, the take-home message of Skyrms's models is that the existence of signals and signalling is not mysterious. Information exchange emerges spontaneously, without the need for elaborate cognitive capacity or pre-existing communicative capacity, over a wide range of circumstances.

2. VERVETS AND HEDGEHOGS

Skyrms's models are both simple and elegant, and his discussion of them is both nuanced and explicit. Moreover, his models are often suggestively juxtaposed with examples drawn from the natural history of signalling; of real agents pursuing individual or collective interests by signalling to one another. A critical question is whether, or to what extent, the models explain the natural history of signalling. Models inevitably simplify, so it's always possible to wonder whether the right simplifications have been chosen.Footnote ¹ For this reason, it is instructive to compare Skyrms's idea that signalling emerges spontaneously, over a wide range of circumstances, with the more guarded message of Godfrey-Smith (Reference Godfrey-Smith1996). Godfrey-Smith considers a third option for the agent who would need to read the signal: pick a response that pays, though not perfectly, in both states of the world. Call this the hedgehog response. If the world is in state1, act1>hedgehog>act2. If the world is in state2, act2>hedgehog>act1. Hedgehog options, when available, may well block the emergence of signalling. That will depend on the costs of learning and the payoff difference between the optimal act for each state and the hedgehog payoff for each state. If learning is expensive or the difference is small, play hedgehog rather than learn to send and read signals.

Let me explain in a little more detail. Godfrey-Smith's target is phenotypic plasticity (the capacity many organisms have of adaptively changing their form or their behaviour in response to a change in their environment), not communication. But making and responding to signals is just a special case of phenotypic plasticity. As I have just noted above, there is an important difference between Godfrey-Smith's approach and that of Skyrms. Skyrms's signal receivers have to make a forced choice: they have at their disposal an act1 that pays off in and only in state1, an act2 that pays in state2, and so on. So they must commit to a hypothesis about the state of the world. Picking a hedgehog response is an alternative to guessing about the state of the world (and thus learning to attend to signals to improve the accuracy of the guess). This action pays reasonably well across all the relevant possibilities, without being optimal in any. The broad availability of such general-purpose responses is an open empirical question. Vervets, famously, tune their anti-predator behaviour to the specific threat they face: they have no generalized response to predation. Hedgehogs do, rolling into a ball with spines erect, whenever they feel threatened. Imperfect, but quite useful against a range of threats (though not against approaching vehicles).

Skyrms rarely considers responses that work reasonably well in both conditions. But if a viable hedgehog strategy is available to an agent, it will pay that agent to learn to tune his behaviour to specific states of the world only if learning is quick and cheap, or if there is a large difference between the return from the optimal strategy and the return from the hedgehog strategy. Signal and response strategies are less likely to emerge, then, in contexts in which a decently paying all-purpose response — hunting the hare — is an alternative to co-ordination. There might often be such cases. That suspicion is reinforced once we recall that in real situations of interaction, no signal is perfectly reliable: even when the agent learns that a bark means ‘leopard’, not all barks will be accompanied by leopards, and not all leopards will be greeted with barks. So attending to signals will never perfectly match responses to states of the world.

So, first caution. Perhaps Skyrms's choice of model structure weighs the dice in favour of the emergence of signals. Second caution: even in vervet-like cases, perhaps the models do not map smoothly onto the phenomena. The simplicity of Skyrms's models make them tractable to formal investigation. But the same simplicity pushes the models and the exemplars apart. Indeed, even the very simplest real-world cases are very different from the models. As we have just noted, vervet alarm calls are a celebrated example from natural history. Dorothy Cheney and Robert Seyfarth made famous the fact that vervets have predator-specific alarm calls, presumably because the best response to threat depends on the type of predator (Cheney and Seyfarth Reference Cheney and Seyfarth1990). Skyrms thinks that the vervet signalling system is a natural example of a Lewis signalling game, with the monkeys having found their way to one of the efficient equilibria (pp. 22–23). So, as Skyrms thinks of it, the world states are eagle-present, leopard-present and snake-present. The actions are hide-in-the-undergrowth; run-up-a-tree; scan-and-move-away. The monkeys have found the following signalling equilibrium:

But while this might look like a Lewis signalling game, it is not one. Vervet alarm calling is not a solution to a coordination problem, for there is no coordination problem. My payoff to running up a tree in response to a leopard threat does not depend on your doing so too. Indeed, I am probably marginally safer if you do not run up a tree: the leopard is likely to attack the vulnerable monkey in the open rather than chase me up a tree. In the Lewis game agents bring different capacities to the table, and their reward depends on coordination. If vervets mobbed predators, it might be more like a Lewis game, but real vervets signal despite not having mutually dependent rewards; that is part of the puzzle of their system. Warning calls are widespread in nature, but that fact is more puzzling than one would suppose from Skyrms's presentation. The vervet example shows that it is often hard to see what is in it for those making the call (hence exotic theories such as in Zahavi and Zahavi Reference Zahavi and Zahavi1997). But it is also hard to see how calls are introduced and stabilized by learning (though clearly they can be, as Skyrms shows through a vivid example on p. 120). The problem is to see how an appropriate response to an alarm call could be coded as a reward. For the animal's experiences are subjectively unpleasant: flight, with fear and stress; perhaps even injury. So the animal hears a signal S, tries act A, and has a most unpleasant time. Act A should become less likely in the presence of S.Footnote ² Of course, the alternative would have been far worse. But how does an automatic, subcognitive, score-keeping system know this? More generally though, the vervets illustrate a pervasive problem for the project. In building the book around simple models, Skyrms is faced with important differences between the models he explores and the phenomena that, presumably, the models are meant to explain.

3. THE STRUCTURE OF COORDINATION

In the last section, I noted a potential problem for Skyrms's intellectual strategy: the inevitable contrast between the simplicity of his models, and the messiness of the real-world systems the models are meant to explain. That said, simple models – models which abstract and idealize – can be explanatorily powerful. Iterated prisoner's dilemmas (and the somewhat more complex elaboration of the basic idea in Skyrms (Reference Skyrms1996)) explain important aspects of cooperation, including human cooperation, despite their brutal simplicity (Binmore Reference Binmore1994, Reference Binmore1998). They do so because they capture causally fundamental features of interactions. Often, agents interact with one another in a context in which there is a partial overlap of interest (mutual defection is good for no-one); in contexts in which interactions are likely to be repeated; and in contexts in which there are potential mechanisms that link others’ future policies with your current choice. Of course, there are many complications, and they are important. We need other tools to capture the importance of partner choice and the complexities of many–many interaction. But overlap of interest, the shadow of the future, and contingency of future acts on current choices are causally fundamental features of cooperation captured by simple models.

The benefits of cooperation do not always involve coordination, and coordination does not always require signalling. Collective action can reduce individual risk and/or increase individual benefit just through sheer numbers. A pack of African wild dogs can bring down large prey by attrition, through numbers and persistence, not requiring coordination and role specialization. Even when coordination is required, as perhaps in chimpanzee monkey hunts,Footnote ³ the chimpanzees do not need to receive and send special signals. They can observe the interacting agents and respond to the moves of both the target and the other hunters. Coordination can be based on cues: features of an agent's phenotype that are present for reasons independent of the agent's social life, but which can be used as evidence by other agents (Hauser Reference Hauser1996; Danchin and Luc-Alain Reference Danchin and Luc-Alain2004). Thus one pied stilt in a flock sees a fox and flies, and all the others follow it into the air, using its escape behaviour as evidence of danger. For Skyrms, a cue is not a signal.Footnote ⁴ For a signal is arbitrary, in the sense that there are alternative, equally efficient systems of world-signal-action loops (pp. 7–9). So Skyrms's project is not cooperation and coordination in general; it is signal-guided coordination. So the question becomes: do coordination dilemmas mediated by signals have structural similarities, and if so, are those similarities well captured by the signal-based models explored in this work? There may be such similarities to capture. Brett Calcott has argued that just as distribution dilemmas have an abstract structure, so too does the generation of benefit. There are general structural features of agents, tasks and interactions that help explain the benefits of cooperation (Calcott Reference Calcott2008). I think Calcott is right, and that Signals offers some insight into these structural features. But its approach is limited in two important ways. First, the analysis of specialization and the division of labour is too abstract to capture constraints on coordination. Second, I am sceptical about the choice of the Lewis signalling game as the base case.

4. INFORMATION VERSUS INFLUENCE

I have saved the most fundamental challenge to the last. In a famous paper, Krebs and Dawkins (Reference Krebs, Dawkins, Davies and Krebs1984) argued that it was a mistake to see biological signalling systems as systems of information flow. Senders and receivers are in competition, and have no interest in sharing information. Instead, we should see signals as attempts by senders to manipulate receivers; receivers in turn will be under selection to resist manipulation. The evolutionary dynamics of signalling systems are those of an arms race, rather than of increasingly efficient coordination. No problem. Skyrms can accommodate conflicts of interest. He shows that signalling can establish and persist in environments so long as there is some overlap of interest, and despite some deception. And real agents do have overlapping interests. The healthy and agile gazelle that stotts to a cheetah sends a signal that is worth attention despite the very divergent interests of predator and prey. A futile pursuit is expensive to both, and so a reliable signal that a chase will be futile is worth decoding. So the Krebs–Dawkins view of signalling does not undermine the project of Signals.

Owren et al. (Reference Owren, Rendall and Ryan2010) have recently restated a sceptical view of signals and information, and their arguments are less easily met. First, they argue that the concept of information is not defined in an empirically tractable way. There is, of course, a crisp formal concept of information and information content, descending from Shannon. As is well known, the basic idea is that a signal carries information to the extent that it reduces previous uncertainty about the state of the world. How much information a signal carries depends on the reliability of the signal, the range of other signals available, and the difference between the pre-signal and post-signal probability of the state of the world. Thus if leopards are very rare and barks are a reliable and precise sign of one being present, a vervet's bark alarm carries a lot of information. Skyrms builds on this concept (even defining a notion of meaning using it, extending some ideas of Dretske). But it is one thing to have a formally well-specified concept of information: another to apply that concept to real systems in ways that allow empirical tests of alternative hypotheses. To apply Skyrms's concept to actual systems, we need a well-defined space of alternative possible states of the environment, together with their standing probabilities. We also need a well-defined space of alternative signals. These presuppositions are far from trivial. For example, in most real animal systems, the modes of signalling – amplification, repetition, position, and associated gesture – are all functionally important. Such call systems are not discrete, and there is no obvious way of specifying the alternative signalling options. Suppose we wanted to contrast a cooperative hypothesis about the function of signalling with a competitive hypothesis (say, a show-off hypothesis). It is by no means clear how we would apply the formal machinery to a real case in order (say) to determine whether an agent was sending as much information as predicted, less, or more. So Owren et al. argue that in practice, those studying signalling systems ignore this formal machinery, relying on pre-theoretic intuitions about the information content of signals.

Their second line of argument puts pressure on Skyrms's idea that signalling systems are systems of arbitrary signs. In responding to the objection that quorum-sensing in bacteria are ‘just chemistry’, Skyrms accepts that biochemistry constrains bacterial signalling systems but argues that there is enough evolutionary contingency in bacterial mechanisms to support the idea that these are signals, not mere brute causes, because there are alternative possible pathways between state of the world, sender's signal and receiver's response. ‘If we look at quorum-sensing over evolutionary time, and reflect on the variety of uses to which the same basic system has been put, we can recover a sense of the plasticity of signalling. Pure convention has gone, but development of the same ancestral signalling system could go one way or the other’ (p. 31). Drawing on a range of examples from sexual signals in frogs and birds, Owren et al. argue that, for the most part, signal choice is highly constrained. It is constrained by channel conditions: thus a male frog calling to attract potential mates must be audible and localizable above the levels of ambient sound in typical environments (with luck, minimizing exposure to local predators). It is constrained as well by the pre-existing sensory capacities and biases of signal targets. And these, in turn, are often largely fixed by a combination of the phylogenetic inheritance of the lineage and by the ecological demands on the female. She must find food and shelter and avoid predators, and these requirements will shape her capacity to detect and respond to male signals. In part, the objection is that the appeal to information is redundant. We can explain the organization and operation of signal systems without having to ask ourselves what the signal is saying. Talk of information gives us little explanatory leverage. In part, it is a denial of contingency. Even over evolutionary time, there is relatively little contingency in animal signal systems.

Owren et al. seem to overstate their case. It is hard to see, say, how vervet phylogeny and sensory physiology force the bark rather than the cough to be the leopard call. Even so, arbitrariness does not seem to capture the explanatory role of the appeal to information. The honeybee system may well be sharply constrained by honeybee physiology and evolutionary history. Even so, the mapping of dance orientation, distance and intensity into food source direction, distance and value surely is indispensible in explaining the evolution and stability of the system. That would be true, even if it turns out that there is no way bees could signal this information except in the way they do signal it. Conventionality and contingency cannot be playing the role Skyrms assigns to them in marking the difference between information flow and mere brute causation.

5. MODEL AND WORLD

Skyrms's project is difficult to evaluate as a whole, because it lies at the intersection of two different debates, one in philosophy of science, the other, biology. First, in philosophy of science there is a developing debate on the nature of ‘model-based science’. So, for example, Peter Godfrey-Smith (Reference Godfrey-Smith2006, Reference Godfrey-Smith2009) has argued that model-based science aims to represent the world indirectly rather than directly. Darwin, in the Origin, was not doing model-based science. He developed a direct representation of the world; that is, he developed detailed descriptions of general features of the biological world. Darwin is taken to be a great scientist because those descriptions were true (or perhaps approximately true), and the mechanisms Darwin described were of great importance. In contrast, in model-based science, the theorist develops a description of a model, an imagined or fictitious system. An evolutionary modeller might, for example, describe an imagined population of organisms that is infinite in size, with random mating, and with synchronous births and deaths, so all organisms are of the same age. They might show that in this population, even very slight fitness differences lead to major changes in the population over time. The model system is informative (one hopes) because there are resemblances between the model and the world. Even so, what counts as a good model is far from obvious.Footnote ⁸ For as Godfrey-Smith points out, model-based science is productive in part because modellers need not specify in advance which aspects of their models are supposed to resemble their real-world targets and in what ways. Rather, their strategy often involves exploring the properties of models, and using features of models as guides to probe target systems in the world. Good models are tools which guide the construction of more direct representations of target systems. Skyrms's work is a paradigm of model-based science, and most of the explicit content of Signals explores the models: there is very little explicit commitment to model-target system resemblance.

Thus one reason why Signals is hard to evaluate is because the criteria for evaluating model-based science are difficult to apply at a time: the fact that a particular family of models is insightful becomes clear, when it does become clear, only over time. But the project also depends on a controversial choice within biology. In somewhat different ways, Ernst Mayr and Nikko Tinbergen proposed a theoretical division of labour (Mayr Reference Mayr1961; Tinbergen Reference Tinbergen1963). Mayr, for example, distinguished between ultimate and proximate explanation. An ultimate explanation of the bee dance language would detail the evolutionary trajectory through which it was assembled, and the selective forces which drove those changes. A proximate explanation would detail the developmental and physiological mechanisms that would implement the dance and its reception. Mayr pointed out that proximate and ultimate explanations complement rather than displace one another, and suggested that (at least to a significant extent) each can be developed independently of the other. Skyrms's project is clearly committed to Mayr's division of labour. With the exception of a discussion on subtle variations in learning rules, he abstracts entirely away from proximate mechanism. That is a risky bet. Everyone accepts that ultimate and proximate explanations are complementary. And most would accept that early in the development of the life sciences, it was reasonable for evolutionary theorists to ‘black box’ developmental biology and physiology (and vice versa). But there is a significant group of evolutionary biologists who argue that this division of labour has outlived its usefulness. Evolutionary developmental biologists, in particular, argue that proximate biology, broadly understood, interacts fundamentally with evolution. We saw this issue in section 3.2, when discussing limits on the evolution of signalling. Skyrms ‘black-boxes’ the organism, and thus bets on the independence of historical trajectories from mechanical implementation. As a consequence of these theoretical choices, while the book is a beautiful exploration of the structure of a family of models, it may not guide our exploration of the natural history of communication and co-ordination.Footnote ⁹