Goldman, A. I. (1992). In Defense of the Simulation Theory. Mind And Language, 7(1-2), 104–119. doi:10.1111/j.1468-0017.1992.tb00200.x

Summary

In this paper, Goldman replies to Stich & Nichols, conceding some of the points they make but also providing further evidence for simulation theory.

Goldman begins by first pointing out that the simulation theory isn’t a substantial departure from previous approaches in cognitive science, as it falls under the umbrella of knowledge-poor approaches (e.g. heuristics and biases) in contrast to knowledge-rich approaches (e.g. rules and symbols). Thus, the simulation theory isn’t a “radical departure” from other paradigms in cognition.

In the next section, Goldman points out that theory theory, as described by Stich & Nichols, depards substantially from other formulations of the theory theory. For example, he states that “in the philosophical literature it has been widely assumed that it should be easy to formulate the principles of folk psychology because they are platitudes, i.e. truths that are obvious to everyone” (pg. 106). Additionally, Goldman discusses “the assumption that folk psychological platitudes are culturally produced and culturally transmitted” (pg. 106-7). Both of these claims seem dubious, and presumably Stich & Nichols woult not adhere to either of them. Thus, their definition of theory theory isn’t the same as everyone else’s. So, while criticisms of claims like those don’t necessarily “knock-down” the theory theory, “these arguments do cast doubt on some popular variations of the theory-theory theme, and highlight the difficulties that must be met by any detailed development of the theory-theory” (pg. 108).

Next, Goldman discusses the relationship to simulations in other domains (i.e., mental imagery). Here he makes the distinction between process-driven and theory-driven simulation, and concedes that mental simulation may indeed be theory-driven; thus, it is necessary to show that the simulation theory is process driven, not just that it is a simulation. This leads into a discussion on introspection, which cannot on its own be used to discriminate between process- and theory-driven simulations. Goldman argues that, for the present purposes, the point of introspection isn’t to distinguish between the two, just to show that theory of mind may indeed involve some form of simulation (as a first step).

Goldman next responds to Stich & Nichols point about simplicity, which was that the theory theory gets the control mechanism “for free” while simulation theory gets the database “for free”. He makes a good point that it isn’t entirely clear what “for free” means, and argues that simulation theory does indeed get the control “for free” by arguing that the process that interprets the output of the decision-making process needs to be present for any theory, and thus it would indeed be available to off-line simulation as well. (I think Goldman misses the point of what “control” means here, but I’ll get back to this in the takeaways section).

Next, Goldman discusses additional evidence from autism for simulation theory. Specifically, that there is evidence that autistic children are perfectly capable at theorizing about mechanistic or behavioral processes (just not mentalistic ones), and that they do have a concept of desire, but that they have difficulty evaulating what is a “reasonable” desire based on the situation. He concludes that these pieces of evidence are easily explained by simulation theory, but not theory theory.

He also discusses the developmental literature discussed by Stich & Nichols, and argues that the results they cite are contradicted by other studies and may be due to confusing task demands. He also suggests the possibility that younger children may understand beliefs, but do not classify those beliefs as being the same as knowledge.

Finally, Goldman replies to Stich & Nichols’ arguments about cognitive penetrability, and in particular, the claim that simulations should predict the same impenetrable behaviors that people exhibit when they make predictions about others. Goldman argues that this only should be the case if the inputs to the simulation are identical to when people are in the situation themselves, and only if the interpretation of those simulations is identical as well. He questions whether this would actually be the case in the examples cited by Stich & Nichols.

Takeaways

I like Goldman’s further characterization of simulation as being either theory-driven or process-driven. That is, the simulation is either driven from a theory (and therefore is something like a simulation of a mental model), or it is actually an execution of a real process in the brain. In the latter case I wonder if it is even really appropriate to call it “simulation”, as it’s not a simulation of the process by which you would act, it is the process by which you would act. A simulation actually implies that it is a copy of a process that necessarily makes certain assumptions and simplifications. I guess it is a simulation in the sense that the real process is being used in the context of a simulation (i.e., pretend inputs).

In discussing the role of the controller, I think Goldman misses the point of why this is important. I made this point in my notes on Gordon as well: knowing what “pretend” inputs to feed the simulation is a nontrivial task. It cannot simply be a matter of trial-and-error. There has to be some additional mechanism guiding the decisions to change certain dimensions of the input rather than others, and what they get changed to. This is related to the difficult problem of determining what information is relevant and what information is irrelevant; knowing the answer to these questions requires some sort of higher-level abstract knowledge (like a theory!).

I do think Gordon brings up some good points regarding empathy. For example, people often feel physiological pain when they see someone else in pain (e.g. if you watch a video of someone cutting their finger with a knife while cutting vegetables, it feels painful to see). Thus, it is not implausible that there is some amount of processing that goes on that engages the same systems that we use when where are in the situation ourselves. However, I don’t think this can be the only explanation. There are also cases where we understand someone else’s behavior while feeling relatively apathetic about it. If I simply read “John cut his finger”, I don’t feel physical pain, but if I see a video of the same thing happening, I do (though, of course, even that is still not as painful as actually experiencing it). But, I still understand how John is feeling in either case. Based on the arguments of Gordon and Goldman, I don’t see how just reading “John cut his finger” would be something that could be fed into the simulator without actually constructing the full scenario (i.e. putting it into a format that the simulation can handle), in which case it ought to predict that you would feel pain.

This whole debate seems so black-and-white, though. Why does it have to be only theory-theory, or only simulation theory? Is it really so impossible for there to be some combination of using higher-order structured knowledge (i.e., a theory) in combination with reuse of existing perceptual, motor, or emotional systems (i.e., simulation)?