Unedited preprint of a Commentary on Anderson & Lebiere "The Newell Test for a theory of cognition" to appear in Behvioral and Brain Sciences
I present reasons for adding an 'embodiment' criterion to the list defended by Anderson and Lebiere. I also entertain a likely objection contending that embodiment is merely a type of 'dynamic behavior', and is therefore covered by the target article. In either case, it turns out that neither connectionism nor ACT-R do particularly well when it comes to embodiment.
The principle that cognitive theories should be evaluated according to multiple criteria is worth adopting, and Anderson and Lebiere's development of Newell's proposals in this regard is useful. One important criterion seems to be missing, though, and that is embodiment.
By embodiment I understand, loosely, physical implementation in an environment. Humans, clearly a key consideration of the target article, are, of course, embodied. They exhibit striking virtuosity at moving around the world, and exploiting the resources available in it. Perhaps more importantly for present purposes, we are talented at exploiting the structure of environments (and of our bodies in them) for cognitive ends, or, as some would have it, engaging in 'environmentally distributed cognition'. One example is locomotion, where recent research (Thelen and Smith 1994) indicates that the architecture of the body, and the properties of the body in interaction with the environment, play significant roles in control of behavior. Another, rather closer to the concerns of traditional cognitive science, is the game of Tetris, where it has been shown (Kirsh and Maglio 1994) that human players use external actions to improve the efficiency (speed, accuracy, error rate) of the spatial manipulations and judgements demanded by the game. External rotation of a Tetris piece, along with inspection to establish whether the rotated piece is in a preferable orientation to before, is often faster and less error-prone than mental rotation for the same purpose. This suggests that at least some cognitive problems are tackled using a coalition of internal and external resources, and that an important feature of our cognitive makeup is that we can detect opportunities for this. (Further examples in humans, other animals and (some) robots abound. Clark 1997 is a useful survey.) This in turn indicates that a theory of cognition that fails to take embodiment seriously is unlikely to capture such features of our own cognitive performance.
A likely objection here notes that Anderson and Lebiere's criterion five is 'dynamic behavior'. Since this criterion concerns the relationship between a cognitive system and an environment perhaps, properly understood, it includes embodiment and distributed cognition. Distributed cognition just is, the objection goes, a kind of dynamical coupling between an information-processing system and a structured body and environment. This objection may be taking charitable interpretation too far. Anderson and Lebiere's discussion of their 'dynamic behavior' criterion (sect. 2.5 of the target article) places considerable emphasis on dealing with the unexpected, and relatively on exploiting external structure. When evaluating the relative performance of classical connectionism and ACT-R with respect to the dynamic behavior criterion (sect. 5.5 of the target article) their emphasis is on real-time control, not embodiment. Rather than try to settle the question whether embodiment is or is not a version of dynamic behavior, I propose to consider how connectionism and ACT-R fare in the case where it is added as a separate criterion, and where dynamic behavior is interpreted to include it.
Were embodiment added as a criterion, I suggest that connectionism would achieve mixed results. In some cases it does extraordinarily well. Consider Quinn and Espenschied's (1993) neural network for controlling a hexapod robot. The success of this system depends to a significant extent on allowing features of the physical construction of the robot, in interaction with the environment, to play a role in control - so that the motion of individual feet will be inhibited if other specific feet do not yet have secure positions. One way of understanding this is to regard the changing physical links between some neurons, parts of the robot anatomy, the physical environment, other parts of the anatomy and (eventually, and sometimes) other neurons, as functioning like additional neurons, or inter-neuron connections, transforming or transmitting information about footing, load on joints, etc. In other cases, though, it is not (yet) clear how to go about building a network, embodied or otherwise, to handle tasks (such as air traffic control) involving fairly specific and detailed functional decomposition, tasks for which systems such as ACT-R seem well suited.
ACT-R, I argue, scores worse for embodiment. Its successes at, for example, modelling driving are in constrained simulation environments, where embodied interaction with the 'feel' of the vehicle and its relation to the road surface, are absent, and where attendant opportunities for exploiting environmental structure (engine tone, vibration) to help cue such actions as gear changes are absent for both the human subjects who provide the target data, and the ACT-R models of driving behavior which do well at approximating the behavior of such humans.
ACT-R, I argue, scores worse for embodiment. Its successes at, for example, modelling driving are in constrained simulation environments, where embodied interaction with the 'feel' of the vehicle and its relation to the road surface, are absent, and where attendant opportunities for exploiting environmental structure (engine tone, vibration) to help cue such actions as gear changes are absent for both the human subjects who provide the target data, and the ACT-R models of driving behavior which do well at approximating the behavior of such humans.
On the other hand, though, we might reinterpret Anderson and Lebiere's 'dynamical behavior' criterion in a way that includes embodiment as a sub-type of dynamic behavior. In this case, and in the light of what is said in the target article and immediately above, connectionism should retain its mixed score. In this case ACT-R should also, I argue, receive a mixed score: It doesn't do well at plain embodiment, but does better at non-embodied forms of dynamic behavior. In either case, the moral to draw is that if embodiment is a genuinely important criterion, then neither connectionism nor ACT-R seem, as they stand, in a good position to perform consistently well on it.
Clark, A. (1997) Being There. MIT Press.
Kirsh, D. & Maglio, P. P. (1994) On Distinguishing Epistemic from Pragmatic Action, Cognitive Science, 18(4), 513-549.
Quinn, R. & Espenschied, K. (1993) Control of a hexapod robot using a biologically inspired neural network. In Biological Neural Networks in Invertebrate Neuroethology and Robotics, ed. R. Beer et al. Academic Press.
Thelen, E. & Smith, L. B. (1994) A Dynamic Systems Approach to the Development of Cognition and Action. MIT Press.