To summarize the status quo, it seems fair to say that, despite its immense importance for the functioning of economic theory, theoretical approaches to learning in economics have so far addressed learning not as a behavioral feature, but – to paraphrase Blume and Easley – as a poorly chosen euphemism for farfetched statistical techniques and passive mechanisms of information accumulation and updating under mostly perfect situational conditions that cannot be justified neither by a sound concept of human learning, nor by empirical success.
Since experiment-based approaches are only a more recent development, it may be premature to judge the results, but so far only few learning mechanisms that would cover a wider class of models (or games) have been proposed, and attempts to test such a mechanism are relatively rare. As the number of models and experiments to test them increases, mechanisms can be expected to improve their ability to capture observed learning behavior, especially when the performance of models is assessed not only through hypothesis testing but through competitive tests among alternative models with respect to the same set of data (BUSH 1963).
At the same time criteria for model comparison have to be developed that are widely missing today and that involve a number of difficult problems.[39] Nevertheless, the issue of model comparison seems crucial for the progress of this branch of research. However, even if the problems of model comparison can be solved successfully, the question remains what results can be expected from this type of research. I suspect that one will find a general model of learning that applies to a variety of situations very well.
Most likely there will be a model that "fits best on average" over a certain range of games or situations, but there will be many more models that fit better to specific situations or games since they are adjusted, modified or calibrated. Even if a single relatively general mechanism can be found, what can we conclude other then that some fundamental principles of learning which are welldocumented in the psychological literature indeed appear to hold in interactive settings?[40] These findings may be interesting per se – though more to psychologists than to economists.[41]
Nonetheless, for most applications in economics, rather than knowing the particular learning algorithm that governs behavior in a specific situation, it may be even more useful to understand the determinants or conditions that influence behavior in a variety of situations, and the limits imposed by these conditions. Hence, questions related to the application of an economic learning theory or its practical implications – for example questions of market design, the occurrence of anomalies, or government interventions in markets (all in which learning may play an important role) – may be difficult to answer within the prevailing methodology.
To address this type of questions more directly, a different approach that parallels the traditional methodology in economics more closely may be appropriate, as will be discussed in the next Section. [For a short comparison of current approaches with the contingent learning approach see the Appendix.]
39 Criteria for model comparison must be stipulated and tradeoffs among them must be discussed. These criteria include, for example, the goodness-of-fit, the number of free parameters (parsimony), the economic interpretability of these parameters, and the validation of parameters in experiments.
40 These principles are likely to include the “Law of Effect” (Thorndike, 1898), the “Power Law of Practice” (Blackburn, 1936), reference points effects, and the effects of negative versus positive reinforcements (see DANIEL, SEALE & RAPOPORT, 1997, 18).
41 The goal of at least some of these learning models is to account for individual, not aggregate behavior. To economists individual behavior has traditionally been of interest only in so far as something can be learned about aggregate outcomes. That is, if individual behavior differs substantially among subjects (and from theory), but aggregate behavior or outcomes are in harmony with the theory (as the results of RAPOPORT ET AL. (1995) suggest), most economists may be reluctant to these findings since economic theory is one of aggregate outcomes, based on average individual behavior.
Prof. Tilman Slembeck
Next: Methodological Issues
Summary: Index