| Yhteenveto: | This thesis is concerned with decision making within an enterprise. It introduces a research framework which is based on the soft computing methods of the control theory: fuzzy logic, neural networks and genetic algorithms together with the signal processing theory. The latter attaches time/frequency dualism to the framework. To prove the validity of the introduced framework it is applied into two financial engineering problems: portfolio and enterprise analysis and into one control engineering problem: fine paper quality optimization. The aim of these applications was to prove the close relationship between financial and control engineering, especially the former in those areas where quantitative methods can be applied. Traditionally in econometrics and enterprise analysis statistical methods are applied. This thesis proposes that the methods of the modem control theory are also applicable. The applicability of the research framework into decision making is proved by developing a quite large expert system. In the scope of this expert system this thesis holds that there need not be very much redundancy when a general expert system is applied into various decision making tasks. In neuro-fuzzy expert systems this is due to the fact that they can extract their production rules from empiric data which differ them from symbolic rule based systems. In the financial engineering part a new neural network algorithm for Fir synapse is introduced and it is compared with some recently proposed methods. Deviating from the traditional gradient descent training method this algorithm initializes the network via genetic algorithms and optimizes the network by conjugate gradient method and simulated annealing. The development of a new algorithm is due to keeping the research framework unified with other methods which are applied in this thesis, especially with fuzzy rule extraction. Also for reinforcement learning a new algorithm is introduced.
|
|---|