Showing papers by "Shun-Feng Su published in 2002"

Robust support vector regression networks for function approximation with outliers

Abstract: Support vector regression (SVR) employs the support vector machine (SVM) to tackle problems of function approximation and regression estimation. SVR has been shown to have good robust properties against noise. When the parameters used in SVR are improperly selected, overfitting phenomena may still occur. However, the selection of various parameters is not straightforward. Besides, in SVR, outliers may also possibly be taken as support vectors. Such an inclusion of outliers in support vectors may lead to seriously overfitting phenomena. In this paper, a novel regression approach, termed as the robust support vector regression (RSVR) network, is proposed to enhance the robust capability of SVR. In the approach, traditional robust learning approaches are employed to improve the learning performance for any selected parameters. From the simulation results, our RSVR can always improve the performance of the learned systems for all cases. Besides, it can be found that even the training lasted for a long period, the testing errors would not go up. In other words, the overfitting phenomenon is indeed suppressed.

A genetic algorithm with domain knowledge for weapon-target assignment problems

Abstract: In this paper, a novel genetic algorithm, including domain specific knowledge into the crossover operator and the local search mechanism for solving weapon‐target assignment (WTA) problems is proposed. The WTA problem is a full assignment of weapons to hostile targets with the objective of minimizing the expected damage value to own‐force assets. It is an NP‐complete problem. In our study, a greedy reformation and a new crossover operator are proposed to improve the search efficiency. The proposed algorithm outperforms its competitors on all test problems.

A high precision global prediction approach based on local prediction approaches

Abstract: Traditional model-free prediction approaches, such as neural networks or fuzzy models use all training data without preference in building their prediction models. Alternately, one may make predictions based only on a set of the most recent data without using other data. Usually, such local prediction schemes may have better performance in predicting time series than global prediction schemes do. However, local prediction schemes only use the most recent information and ignore information bearing on far away data. As a result, the accuracy of local prediction schemes may be limited. In this paper a novel prediction approach, termed the Markov-Fourier gray model (MFGM), is proposed. The approach builds a gray model from a set of the most recent data and a Fourier series is used to fit the residuals produced by this gray model. Then, the Markov matrices are employed to encode possible global information generated also by the residuals. It is evident that MFGM can provide the best performance among existing prediction schemes. Besides, we also implemented a short-term MFGM approach, in which the Markov matrices only recorded information for a period of time instead of all data. The predictions using MFGM again are more accurate than those using short-term MFGM. Thus, it is concluded that the global information encoded in the Markov matrices indeed can provide useful information for predictions.

On the dynamical modeling with neural fuzzy networks

Abstract: In the literature, researchers have introduced delay feedback (or recurrent) networks and claimed that those networks could accurately model dynamical systems without knowing their system orders. In this paper, we have studied those delay feedback networks and also proposed a better version of delay feedback neural-fuzzy networks, called additive delay feedback neural-fuzzy networks (ADFNFN). From our simulations for various examples, it is clearly evident that ADFNFN can have the best modeling accuracy among those existing delay feedback networks. Nevertheless, we also showed by examples that those delay feedback networks can only reach the accuracy of nonlinear autoregressive with exogenous inputs (NARX) models with order two, and that the number of delays in delay feedback networks plays the same role as the order in NARX models.

Robust CMAC control schemes for dynamic trajectory following

Abstract: Improved robust CMAC control schemes are proposed for tracing dynamic trajectories in this paper. There are two main structures in the proposed control schemes: one is the robust controller and the other is the improved CMAC network. The robust controller technique can achieve a certain goal without concern for instability of the controlled system in the presence of significant plant uncertainties if the nominal parameter is roughly estimated. Next, in order to reduce the tracing error, a suitable nominal parameter needs to be chosen. Thus, the improved CMAC learning approach under the robust control structure, using the concept of credit assignment, will be employed to determine control variables that can trace other states repeatedly during control processes. Finally, simulation results demonstrate the capability of the proposed control schemes to trace dynamic trajectories.

Embedding fuzzy mechanisms and knowledge in box-type reinforcement learning controllers

Abstract: In this paper, we report our study on embedding fuzzy mechanisms and knowledge into box-type reinforcement learning controllers. One previous approach for incorporating fuzzy mechanisms can only achieve one successful run out of nine tests compared to eight successful runs in a nonfuzzy learning control scheme. After analysis, the credit assignment problem and the weighting domination problem are identified. Furthermore, the use of fuzzy mechanisms in temporal difference seems to play a negative factor. Modifications to overcome those problems are proposed. Furthermore, several remedies are employed in that approach. The effects of those remedies applied to our learning scheme are presented and possible variations are also studied. Finally, the issue of incorporating knowledge into reinforcement learning systems is studied. From our simulations, it is concluded that the use of knowledge for the control network can provide good learning results, but the use of knowledge for the evaluation network alone seems unable to provide any significant advantages. Furthermore, we also employ Makarovic's (1988) rules as the knowledge for the initial setting of the control network. In our study, the rules are separated into four groups to avoid the ordering problem.