Procedural rationality

5 Conclusion

A necessary but not sufficient condition for simulation models to converge towards a rational expectations equilibrium is that agent perceptions converge so that mean expectations are close to model outcomes. The Sargent research programme has as its objective the demonstration that non-homogeneous agents who search the whole possibility space will converge on perceptions which are effectively consistent with the outputs of the system model. The transition models described above generate convergence in behaviour -- specifically in price-setting behaviour -- by local exploitation of declarative knowledge. These models were not implemented with rational expectations equilibria in mind. They were implemented to demonstrate that models which are well validated with respect to both formal logics and independent theoretical structures relating to learning and decision-making also lead to verifiably accurate model outputs.

I conjecture that any cognitive behaviour which entails the convergence of individuals' perceptions and where their perceptions are changed when systematically wrong will also entail convergence towards expectations which are not wrong in any biased way. If so, then ultimate convergence towards a rational-expectations equilibrium or something similar is a weak condition in the absence of structural change in the economy. However, the process of learning by global search and the process of learning by local exploitation might well yield very different results in relation to the effects of economic and social policies which are intended to effect particular structural changes. Before implementing policies based on one of these approaches or the other, it would make sense to look for reasons to believe in one or the other. Validation with respect for formalisms -- mathematics or predicate or propositional logics -- ensures rigour in the sense of a lack of ambiguity in specifications and implementations of models. Validation with respect to independent, experimentally or otherwise verifiable theories gives us confidence that our own models are better than just data-mining. I can see no reason to have confidence in the encoding of human cognition as genetic algorithms simply because the resulting models converge to something which is not too different from conventional rational-expectations-type results. This scepticism stems from the absence of any independent reason to believe that agents learn by global search and the plethora of evidence that they learn by local exploitation of declarative knowledge conditioned by their procedural knowledge. Since procedural knowledge can only be acquired by experience, the identification and use of declarative knowledge must always be based on the experience of what the agent has done and this is itself an inherently local processing of declarative knowledge.

Experience indicates that economists are unlikely to be influenced by the very different style of process-centred disciplines such as cognitive science. Nonetheless, we have shown the feasibility of implementing models using concepts which have arisen independently in analyses of business history, the economic history of technical change and in the cognitive sciences to explain with empirical verification how learning and decision-making actually take place.

Procedural rationality - 09 DEC 97

[Next] [Previous] [Top] [Contents]

Generated with CERN WebMaker