I am sometimes calling into question the idea of skill acquisition in its dominant representationalist view. In AI the idea of skills has been instantiated under many forms: from the macro-actions in STRIPS-like frameworks, to the behaviors of the subsumption achitecture, or any of the various formalisms of hierachical reinforcement learning proposed during the 90s.

For example, in the options framework a skill is a triple of the form \(\langle \mathcal{I}, \pi, \mathcal{B} \rangle\): an initiation set (precondition), policy and termination function (postcondition). From a practical point of view, this is a convenient way of explaining what it means to reason over temporally extended actions.

But does it even makes sense to treat skills as representable atomic objects of the agent-environement interaction ? Should the rational behavior of an agent be decomposable into elementary skills to truly qualify as a skillfull behavior ?

Calvin (2000) asks (p.112):

So where in all this does my memory of a lighthouse reside ? Or my plan to go there ? Or my one-foot-in-front-of-the-other subprogram ?

Hinting at an answer, he quotes the mathematician Stanislaw Ulam:

[What] you are describing is not an object but a function, a role that is inextricably tied to some context. Take away the context and the meaning also disappears...

Ulam goes on arguing that what we perceive is a function but not an object per se. Araujo and Davids (2011) adopt a similar view. They claim that cognitive and experimental psychology are going through the same phases as that of other sciences which evolved from the study of the substance to that of the inherent process. In physics for example, the substance of caloric heat has eventually been replaced by the process of random kinectic motion. They argue that such a position necessarily leads to an explanatory regress, presupposing the existence of a behavior by some pre-existing structures. The fundamental problem remains to explain why such apparent structures might appear: "a quest for understanding of causation" (Calvin 2000) (p. 124).

Talking about the class of symbolic models of cognition such as ACT-R (Newell 1990), Araujo and Davids (2011) explains that:

The dominant tendency in deterministic system theorizing is to equate knowledge with the acquisition and storage of discrete mental concepts and to study the form of these entities along with the inferential processes (required for symbol manipulation) that operate upon them.

The tendency towards explaining behavior in terms of personal features rather than the agent-environment relation is typical of the representationalist models: a problem of "organismic asymmetry". The ecological view of skill acquisition attributes the emergence of behavior both in terms of personal and environmental constraints. For Araujo, the process of skill acquisition should rather be described as "skill attenuement" achived by attending to the key properties (affordances) of the "performance environment" according to the the agent's body and action set.

References