Complexity and simplicity in the evolution of decision-making biases.

Marshall et al. [1] critique recent evolutionary explanation of decision-making biases, focusing on Johnson and Fow ler’s model of overconfidence [2] and Trivers’ model of self deception [3]. We agree with Marshall et al.’s central prem ise that a Bayesian decision-maker would also be able t optimize fitness in these settings (see their Box 2). Howev er, as their Figure 1 makes clear, the point is that Baye sians require an extraordinary amount of information, a well as processing power, to achieve the same behavior. In the scenario that we modeled, actors were in competition over resources, fighting if both made a claim to a resource To calculate the ‘estimated [fitness] value of claiming’ a resource, a Bayesian decision-maker would have to know (and process) not only: (i) costs and (ii) benefits, but also (iii the prior probabilities of conflict payoffs, and (iv) the prob ability of winning, which is (v) a function of one’s own, and one’s opponent’s strength. Moreover, as Marshall et al themselves note, these variables would also have to b ‘reliably estimated’; information might be available, bu not accurately transmitted or perceived. In our model [2] individuals only require information about a single one o these five parameters, (v) their own and their opponent’ strength. Not only is it simpler, but also it is readily gauged in nature, for example by size differences. As noted above such measurements may be inaccurate, but measuring one thing is easier than measuring five things. The point of the Johnson and Fowler model was not that Bayesian decision making is inefficient, but that organisms and natural se lection would not need such complex decision-making algo rithms to achieve optimal behavior. Perhaps even more problematic is the assumption by Marshall et al. that a decision-theoretic approach is suffi cient to characterize Bayesian decision-making; thei actors do not take into account the possibility that they are interacting with non-Bayesians, in which case thei estimates of probabilities resulting from strategic interac tion will be incorrect. In essence, Bayesians must not only have a model of ‘decisions’ in their head, but also a model o opposing ‘decision-makers’, which depends on information about how those opponents are in turn coping with strate gic interaction and possible information constraints. By contrast, the equilibrium approach in the Johnson and Fowler model [2] shows that individuals with a simple heuristic who have biased perceptions of their own capa bilities are likely to survive when they interact with ‘any