Evolution of a predictive internal model in an embodied and situated agent

Abstract

We show how simulated robots evolved for the ability to display a context-dependent periodic behavior can spontaneously develop an internal model and rely on it to fulfill their task when sensory stimulation is temporarily unavailable. The analysis of some of the best evolved agents indicates that their internal model operates by anticipating sensory stimuli. More precisely, it anticipates functional properties of the next sensory state rather than the exact state that sensors will assume. The characteristics of the states that are anticipated and of the sensorimotor rules that determine how the agents react to the experienced states, however, ensure that they produce very similar behaviour during normal and blind phases in which sensory stimulation is available or is self-generated by the agent, respectively. Agents' internal models also ensure an effective transition during the phases in which agents' internal dynamics is decoupled and re-coupled with the sensorimotor flow. Our results suggest that internal models might have arisen for behavioral reasons, and successively exapted for other cognitive functions. Moreover, the obtained results suggest that self-generated internal states should not necessarily match in detail the corresponding sensory states and might rather encode more abstract and motor-oriented information.

Simulations

“…The thesis has also further described the neural mechanisms underlying simulations. Pezzulo et al (2011) argued more generally that grounded or embodied cognition need to develop computational models that can drive embodied cognition theories to be more detailed such that they can lead to more testable predictions.…”

“…being an epiphenomenon). Pezzulo et al (2011), however, argued that evidence for the ability of sensorimotor processes to modulate the content of abstract conceptual representations, that is, that there are bidirectional synaptic pathways that allow both higher-level areas to modulate or reconstruct activity in lower-level areas and lower-level signals to modulate or activate higherlevel areas (cf. Meyer and Damasio, 2009).…”

“…The first and perhaps more common usage refer to conceptualization (e.g. Barsalou, 2009;Pezzulo, et al, 2011), where simulations are thought to explain also the abstract nature of parts of cognition, that is, how it is possible to abstract generalities among several objects and situations. For example, Barsalou have described a number of different theoretical constructs, such as perceptual symbols and situated conceptualizations to explain this process, as well as, conducting and reviewing the empirical evidence supporting these claims (see e.g.…”

“…Barsalou, 2009;Lakoff & Johnsson, 1999;Pezzulo, et al, 2011). On the other hand, the phenomenological and informational content of the simulations could also depend on the level of the sensory and motor hierarchies at which simulations occur.…”

“…A more detailed model could possibly be more directly compared to behavioral and neuroscientific data (cf. Pezzulo, et al, 2011) Another open issue that has not been addressed explicitly in this thesis is how the brain can switch between simulating an action or perception and acting and perceiving (cf. Svensson, 2007).…”