Resource-rationality models

The previous problematic considerations of RA models led to the development of the so-called “Resource-Rationality” (RR) paradigm (a sort of evolution of the RA paradigm), which tries to incorporate additional constraints with regard to those minimally hypothesized by RA models. Such constraints also consider, for example, which cognitive operations are actually available to a cognitive agent, as well as their time and cost demands. In some senses, the RR paradigm allows the development of “bounded optimal” models (where “optimality” refers to the evolutionary account given by Anderson) by incorporating additional cognitive constraints other than those shaped by evolution (or, using terminology introduced in the previous chapter, we could say that such a paradigm proposes the adoption of more “structural” constraints). The overall idea is that such “bounded optimal” models can then be more easily translated in computer simulations since computer scientists and AI researchers have already developed a theory of rationality that accounts for limited computational resources (Horvitz,

Cooper, and Heckerman, 1989; Russell and Subramanian, 1995). In particular, “bounded optimality” is a theory for designing optimal programs for agents with performance-limited hardware that must interact with their environments in real time. In this setting, “a program is bounded-optimal for a given architecture if it enables that architecture to perform as well as or better than any other program the architecture could execute instead” (Lieder and Griffiths, 2019).

Within the RR theory, the kind of constraints considered are described as follows by Lieder and Griffiths (2019: 15):

Our theory assumes that the cognitive limitations inherent in the biologically feasible minds include a limited set of elementary operations (e.g., counting and memory recall are available but applying Bayes’ theorem is not), limited processing speed (each operation takes a certain amount of time), and potentially other constraints, such as limited working memory. Critically, the world state is constantly changing as the mind deliberates. Thus, performing well requires the bounded optimal mind to not only generate good decisions but to do so quickly. Since each cognitive operation takes time, bounded optimality often requires computational frugality.

In other words, differing from the RA hypothesis, the guiding premise of RR is that resource limitations must be considered as built-in features from the start. The modelling approach for building RR models is summarized in the following five steps:

• 1 Start with a computational-level (i.e., functional) description of an aspect of cognition formulated as a problem and its solution.
• 2 Posit which class of algorithms the mind’s computational architecture might use to approximately solve this problem, a cost in computational resources used by these algorithms, and the utility of more accurately approximating the correct solution.
• 3 Find the algorithm in this class that optimally trades off resources and approximation accuracy.
• 4 Evaluate the predictions of the resulting rational process model against empirical data.
• 5 Refine the computational-level theory (Step 1) or assumed computational architecture and its constraints (Step 2) to address significant discrepancies, derive a refined resource-rational model, and then reiterate or stop. (Lieder and Griffiths, 2019: 21)

So, the starting point is the computational level of analysis introduced by Marr (1982) that we discussed in the previous chapter. Once the function of the studied cognitive capacity has been individuated, the RR paradigm suggests postulating, in functional terms, an abstract computational architecture.

Next (Steps 2-3), the RR approach asks for an analysis of the possible types of algorithms that can optimally solve the problem identified at the computational level, thereby pushing the principles of rational analysis toward Marr’s algorithmic level. Once the model is ready and implemented, its predictions are tested against empirical data. The results can be used to both refine the theory’s assumptions about the computational architecture and the problem to be solved.

This refinement step can lead to a reiteration of the whole RR procedure (starting from Step 1) and can be repeated until the derived model becomes more structurally accurate.

As we saw in the previous chapter, the redefinition of a model’s assumptions may include moving from an abstract computational architecture to increasingly constrained models of the mind or brain (depending on the modelling focus and paradigms used).

The (eventually increased) accuracy of the updated models is indirectly observed via the progressive alignments (if any) between the model predictions and the empirical data. Therefore, the underlying idea is that the results obtained via the implemented system are predictors of how the RR model is increasingly (or decreasingly) closer to the neuro-physical or psychological mechanisms determining human’s responses.

The overall RR process terminates when either the model’s predictions are accurate enough or all relevant cognitive constraints have been incorporated sensibly (Griffiths, Lieder, and Goodman, 2015).

Theories such as the RR, also called “optimization under constraints” theories, have been criticized by competing research programmes for various reasons. One raised objection, similar to the one against the RA approach, is that including constraints in the optimization problem does not make optimization feasible; instead, it makes it harder, so the problem remains intractable. As Gigerenzer, Hertwig, and Pachur (2011: xx) wrote, optimization “becomes more demanding mathematically with each constraint added”. Gigerenzer, in particular, is the main proponent of an alternative theory of human decision making known as Adaptive Toolbox (AT). In this framework, heuristics have a crucial importance since they are considered the basic cognitive tools that compose the “adaptive toolbox” of intelligent living organisms (Gigerenzer, 2000; Gigerenzer, Hertwig, and Pachur, 2011). Interestingly enough, in this framework, as in the original RA accounts of Anderson, some classical decision-making errors and biases are considered powerful - yet fallible - heuristics to make decision in uncertain environments. The AT theory is then designed to avoid the intractability problems ascribed to the CR, RA, and RR accounts, since it assumes that intelligent agents adopt very simple decision strategies that are successful only because they are contextually appropriate. Of course, the questions about “how ‘ecologically rational’ agents are” and “how well any particular heuristic does in any particular environment” are empirical ones and can only be answered via experiments (Goldstein and Gigerenzer, 2002).

It is important to point out, however, that the AT’s and RR’s methodologies are antithetical since the latter requires considering strong rationality assumptions (and constraints) to derive empirical hypotheses; AT, on the other hand, asks about rationality only once the descriptive facts have been laid out. In addition, the AT theory does not consider relevant the appeals to optimization, since it argues that cognition is not optimized and does not necessarily perform optimally in any sense (Rich et al., 2020).

As with RR and RA, however, a critique of the AT theory is that determining in any given case which heuristic to select and apply from our cognitive repertoire is perhaps, again, a very difficult decision problem (Wallin and Gardenfors, 2000) that, from a computational point of view, can lead, again, to intractable solutions. In order to cope with this, RR theorists have proposed including the AT framework in their theoretical apparatus by suggesting that potentially intractable resource-rational decision problems could have been approximately solved by resorting to some of the heuristics proposed by the AT theory (thus combining the two approaches). This has required them postulating an additional (resource-rational) process that yields such heuristics (Lieder and Griffiths, 2019) consisting of a simple meta-heuristic for selecting the AT heuristics. The claim is that the selected heuristics, usually shaped by evolution and simply activated on requests, generally are those leading to decent, “good enough” decisions.

This solution seems to combine the many different modelling and analytics proposals: RA, RR, and AT. It is important, however, to point out that the role of heuristics - and of the other mentioned components - was already emphasized by Simon. As early as 1955, indeed, he seemed to propose a way to put together the most relevant pieces of the theories that would be developed in the following years. Simon (1955), in fact, argued that “rational decision strategies” (i.e., the “satisficing" ones, in his terminology) are those adapted to both the structure of the environment (a crucial element of the RA) and to the mind’s cognitive limitations (the main focus of RR paradigms). He additionally suggested that the pressure for adaptation makes it rational to use heuristics (the key point of the AT paradigm) that select the first option that is “good enough” instead of trying to find the ideal option. In other words, Simon’s ideas inspired each of the three main framework developed in the cognitive modelling community and their integration in search of feasible structural models of cognition.