More attention has been devoted to providing evolutionary accounts of the development of beliefs, or belief-like states, than for desires or preferences. Here I articulate and defend an evolutionary rationale for the development of psychologically real preference states. Preferences token or represent the expected values available actions given discriminated states of world and agent. The argument is an application of the ‘environmental complexity thesis’ found in Godfrey-Smith and Sterelny, although my conclusions differ from Sterelny’s. I argue that tokening expected utilities (...) can, under specified general conditions, be a powerful design solution to the problem of efficiently allocating the capacities of an agent. Preferences are for efficient action selection, and count as a ‘fuel for success’ in the sense urged by Godfrey-Smith for true beliefs. They will tend to be favoured by selection when environments are complex in ways that matter to an organism, and when organisms have large behavioural repertoires with heterogeneous returns and costs.The rationale suggested here is conditional, especially on contingencies in what design options are available to selection and on trade-offs associated with the costs of generating and processing representations of value. The unqualified efficiency rationale for preferences suggests that organisms should represent expected utilities in a comprehensive and consistent way, but none of them do. In the final stages of the paper I consider some of the ways design trade-offs compromise the implementation of preferences in organisms that have them. (shrink)

[A slightly revised version of this paper has been accepted by the BJPS] More attention has been devoted to providing evolutionary scenarios accounting for the development of beliefs, or belief-like states, than for desires or preferences. Here I articulate and defend an evolutionary rationale for the development of psychologically real preference states. Preferences token or represent the expected values of discriminated states, available actions, or action-state pairings. The argument is an application the ‘environmental complexity thesis’ found in Godfrey-Smith and Sterelny, (...) although my conclusions differ from Sterelny’s. I argue that tokening expected utilities can, under specified general conditions, be a powerful design solution to the problem of allocating the capacities of an agent in an efficient way. Preferences are for efficient action selection, and are a ‘fuel for success’ in the sense urged by Godfrey-Smith for true beliefs. They will tend to be favoured by selection when environments are complex in ways that matter to an organism, and when organisms have rich behavioural repertoires with heterogenous returns and costs. The rationale suggested here is conditional, especially on contingencies in what design options are available to selection and on trade-offs associated with the costs of generating and processing representations of value. The unqualified efficiency rationale for preferences suggests that organisms should represent expected utilities in a comprehensive and consistent way, but none of them do. In the final stages of the paper I consider some of the ways in which design trade-offs compromise the implementation of preferences in organisms that have them. (shrink)

There is an enduring tension in thinking about the architecture of systems that select behaviours, including evolved organisms. One line of reasoning supports convergence in control systems and conversion of the values of all options into a common currency, in part because this seems the best or only way of trading off costs and benefits associated with outcomes of varying types. A competing consideration supports parallelism or other forms of fragmentation, because of inefficiencies associated with integration, and suspicion towards general-purpose (...) cognitive systems. In addition, recent neuroscience has given particular attention to the question of how partly independent ‘habit based’ and ‘planning based’ decision-systems might interact. The issues here are, at least in part, a special case of the more general integration versus parallelism tension. This tension provides a useful framework for thinking about the implementation of different kinds of sensitivity to risks of various types, the variety of mechanisms by means of which risk sensitivity might be modulated, and how risk might be traded off against other components of value or disincentive. I describe the tension, briefly survey some of the relevant empirical evidence, and conclude with tentative observations on the topic of risk. (shrink)