Instead of assuming a black box process, some more recent approaches have tried to explicitly model behavior adjustment by introducing statistical techniques and other mechanisms of information updating or gathering, referred to as “learning”.
Three main fields of research can be distinguished. The first addresses learning in (rational) equilibrium models, and the second focuses on learning in a game theoretical context.[13] While these two fields approach learning mainly on theoretical grounds, a third field studies actual learning behavior in experiments. In experimental approaches to learning, researchers aim to find algorithms that mimic actual learning behavior (as observable in experiments) in order to predict that behavior in certain well-defined situations (e.g., in classes of games, see EREV & ROTH, forthcoming). In both theoretical fields researchers face the same problem that many models tend to have multiple equilibria.[14]
Thus, some “learning” mechanism is introduced in order to model adjustment paths that allow to determine or to distinguish between equilibria. The main interest lies in the analysis of the properties under which behavior converges to some equilibrium, and the stability of the equilibrium (if any is reached).
The properties analyzed typically include the rationality assumption, informational assumptions, assumptions about (prior) beliefs, and, of course, the adjustment process itself. Hence, an example of a typical question is: “Does xlearning of w-rational players with y-beliefs (or y-expectations) lead to z-equilibrium?”[15] With focus on the assumed “learning” process (i.e., on x) various mechanisms are discussed in the literature of which the most common ones will shortly be presented in order to characterize the underlying concept of learning.[16]
Learning and Rational Expectations
In the rational expectations literature (where z = rational expectations), the most common adjustment mechanism is Bayesian (or rational[17]) learning. Hence, this and other statistical methods[18] are viewed as representing the human learning processes in the formation of (rational) expectations. This way of modeling may be justified by an “as if” assumption in conjunction with empirical success that would override the involved superrationality (which is assumed in that huge cognitive abilities are needed to perform the calculations).
Unfortunately, econometric evidence seems not to be overwhelming, and the introduction of learning mechanisms does not appear to reduce the set of potential outcomes in a meaningful way.[19] The difficulties lie both in the definition of optimal learning, and in limiting it to one mechanism (out of a large set of plausible mechanisms) that can be justified by some optimality argument.
Also, common statistical techniques cannot be applied in a strict sense to the problem of individual inference in rational expectations (macroeconomic) models since to do so would require individuals to be unaware of the effects of beliefs on outcomes, whereas in these models beliefs affect outcomes and outcomes affect beliefs. Assuming that individuals are unaware of the effects of beliefs on outcomes, however, is unsatisfactory since this would imply that individuals ignore relevant and potentially useful information.
13 Another field may be distinguished that focuses on “social learning” and analyzes “the process by which certain mechanisms in society aggregate the information of individuals”(VIVES, 1996, 589). It draws from both, rational expectations and game theoretic models but has developed genuine concepts (e.g., “herding”, “informational cascades”).
14 Hence, the recent concern with so-called “learning” is not mainly a reaction to the mentioned criticism, but originates from the need for further theoretical development within the prevailing paradigm.
15 Further questions may concern the speed of convergence, the modeling of short- or medium-term behavior (or play), and assumptions about knowledge or information.
16 See KIRMAN & SALMON (1995) for a general overview of current research on learning in economic theory, and FUDENBERG & LEVINE (1997) for a comprehensive work with focus on learning in games.
17 BLUME & EASLEY (1995, 13) refer to rational learning as a “a poorly chosen euphemism for ‘Bayesian learning’.“
18 Following BULLARD (1991) at least three more approaches within rational expectations macroeconomics may be distinguished with respect to the forecast functions employed: forecast functions that (1) use only historical data, that (2) include the beliefs of others, and that (3) include frivolous variables (or “sunspots”).
19 See e.g., SALMON (1995, 236) and BULLARD (1991, 57).
Prof. Tilman Slembeck
Next: Learning in Games
Summary: Index