Summary: | 碩士 === 國立成功大學 === 資訊及電子工程研究所 === 83 === This thesis proposes a reinforcement learning algorithm for
fuzzy neural network with distributed prediction scheme (RFNN-
DPS). The proposed RFNN-DPS model is constructed with a
multilayered fuzzy neural network model. It doesn't require
another neural network predictor for predicting the external
reinforcement signal, and the internal reinforcement
information is distributed into every fuzzy rule. In other
words, the proposed model uses only a network to construct the
system and equips itself with the functions of the action
network and the prediction network. The credit vector of rules
can predict the external reinforcement signal and provide a
more profitable internal reinforcement signal to themselves.
The RFNN-DPS performs a stochastic exploratory algorithm to
adapt itself according to the internal reinforcement signal.
The number of rules can be adjusted dynamically, and the
membership functions of input/output variables can be learned
automatically and incrementally. The RFNN-DPS model operates on
an on-line learning mode, and two algorithms which are the
single-step reinforcement learning algorithm and the multi-step
reinforcement learning algorithm. After the proposed model is
trained, a network structuralized to represent a set of fuzzy
rules retrieved in the reinforcement learning process can be
achieved. In summary, RFNN-DPS model has the advantages of
single model structure, fast learning speed, retrieving the
reliability of rules, and simple learning algorithm.
|