Affordance and action selection

Friston and colleagues, in a 2012 paper on dopamine, recall that Gibson defined affordances as action possibilities latent in the environment. On this definition they are objectively measurable, independent of the ability to recognise them, and always in relation to the actor². The authors add that affordance is an attribute of a cue and has to be inferred. Their paper concerns set switching and action selection during cued responses³. In that account the role of dopamine is to report the precision or salience of cues that portend a predictable sequence of sensorimotor events³.

The same paper simulates the effects of depleting dopamine, including the perseveration characteristic of Parkinson's disease, using synthetic neuronal models based on active inference³. A further result concerns overconfidence about affordance. Such overconfidence lets the second level of the generative model recognise a change in context more quickly. The authors conclude that this produces a trade-off between how efficiently cues elicit movements and how efficiently the set is switched⁴.

Friston, Daunizeau and Kiebel (2009) argued that behaviour can be optimised without invoking reward, value or utility. They solved a benchmark dynamic-programming problem by active inference under the free-energy principle¹. Friston and Ao (2012) made a related point. In their view agents do not seek out valuable states, since valuable states are simply those the agent frequents⁵. They still allowed that minimising expected future cost is a compelling heuristic, which the brain may exploit at higher levels⁵.

Ramstead and colleagues (2019) extended affordance to the niche. They described it as an attribute of active, if counterfactual, engagement with the environment, and also as a statement about the learned, lived world⁸. Reinforcement-learning language has been applied to simpler systems. Whiting and colleagues saw model slime-mould networks aggregate at positively stimulated nodes and suggested this might correspond to positive reinforcement learning⁶. Watson and colleagues compared selection acting on heritable variation in interactions with reinforcement learning, in contrast to unsupervised correlation learning⁷. Doctor and colleagues describe a reinforcement learning agent as learning through trial and error⁹.