There are several varieties of philosophy of biology paper. Some use philosophical analysis to sharpen our understanding of biological theories. Some seek to address general questions in scientific methodology using examples from biology. And some endeavour to apply the formal tools of evolutionary biology and game theory to traditional philosophical questions.
This last type of paper currently represents something of a movement in the philosophy of science. The background commitment is that we, like other organisms, are the product of various evolutionary processes and nothing else. Whatever cognitive capacities we exhibit, then, are ultimately appropriate targets for evolutionary explanation, and the preconditions for the possibility of such capacities are just the preconditions for their evolution.
While one would like to know how our cognitive capacities in fact evolved, that is typically difficult to determine. The fall-back strategy is to explore sufficient conditions for the evolution of our various cognitive capacities. And such how-possible evolutionary stories are more compelling if they proceed from limited, generic resources.
As an example of such a story, in his book Signals Evolution, Learning, and Information Brian Skyrms has given an evolutionary account for how signals with conventional meanings might evolve in a David Lewis signalling game. We will consider a version of this account in the context of simple reinforcement learning.
Suppose that the British have planned a series of invasions. Suppose further that some will be by land and some by sea and that the type of each invasion is randomly determined with equal likelihood. Robert Newman, sexton of the Old North Church, wants to warn the colonists which type of invasion is coming next. He can see which type of invasion is being staged, but he has no meaningful signals that might be used to communicate this information to the colonists.
The sexton does have two urns, one decorated with a picture of land and the other with a picture of the sea, and each urn contains one ball with a picture of a single lantern on it and one ball with a picture of two lanterns. When he sees the invasion type, he shakes the corresponding urn, draws a random ball, then hangs one or two lanterns in the church’s belfry as indicated by the picture on that ball.
The colonists in Charlestown also have two urns, one decorated with a picture of one lantern, the other with a picture of two lanterns, and both containing one ball with a picture of land on it and another with a picture of the sea. When they see one lantern or two lanterns in the church, the colonists shake the corresponding urn, draw a random ball, then prepare for the invasion type indicated by the picture on that ball.
The sexton’s and colonists’ actions on the first invasion are entirely random. The sexton has an even chance of hanging one or two lanterns regardless of the observed actions of the British, and the colonists have an even chance of preparing for land or sea whatever the number of lanterns they see in the church’s belfry.
If the sexton and colonists do not update the contents of their urns, they will continue to act randomly, and no meaningful signals will evolve. But if they update the contents of their urns based on what happens when they signal and act, they might evolve signals that communicate information concerning the nature of the next invasion. Basic reinforcement learning represents one of the simplest ways they might do so.
Suppose that when the agents are successful, that is, when sexton puts lanterns in the belfry and the colonists prepare for the type of invasion in fact being staged by the British, the sexton and the colonists each replace the ball they drew to the urn from which they drew it and add another ball to that urn of the same type; otherwise, each just returns the ball they drew. If, for example, the British are invading by land and the sexton happens to draw a two-lantern ball from his land urn and the colonists happen to draw a land ball from their two-lantern urn and hence successfully prepare for the invasion, then the sexton returns the two-lantern ball and to his land urn and adds a copy and the colonists return their land ball to their two-lantern urn and add a copy.
With such updating of their dispositions, while the agents’ behaviour is initially random, one can prove that for this particular type of evolutionary game (basic reinforcement learning with two unbiased states of nature, two signals, and two actions) the agents will always evolve a signalling system where the sexton consistently sends one signal for land and the other for sea and the colonists always successfully prepare for the corresponding type of invasion. This game then provides an account of how it is possible for very simple, but meaningful, signals to evolve by basic reinforcement learning.
One of the virtues of the concrete evolutionary model is that it is perfectly clear what it means here for the evolved signals to be meaningful. They are meaningful precisely insofar as the evolved signalling dispositions of the agents lead to successful coordinated action.
On reflection, it should be unsurprising that meaningful signals might evolve in the context of a simple evolutionary game. After all, even single-cell organisms coordinate their interactions by way of evolved signals that transmit information. And we take their simple dispositions to be the product of evolutionary processes and nothing more.
One might object that human language is both more complex and more subtle. But here this just means that one needs a correspondingly more sophisticated evolutionary model to account for it. Insofar as we take our cognitive capacities to be the product of evolutionary processes and nothing more, such accounts are there to be found.
