|Budget Amount *help
¥2,800,000 (Direct Cost : ¥2,800,000)
Fiscal Year 2000 : ¥700,000 (Direct Cost : ¥700,000)
Fiscal Year 1999 : ¥800,000 (Direct Cost : ¥800,000)
Fiscal Year 1998 : ¥1,300,000 (Direct Cost : ¥1,300,000)
One of main features in financial investment problems is that the situation changes very often over time. In applying machine learning techniques under this circumstance, in particular, it has been observed that additional learning plays an effective role. However, since the rule for classification becomes more and more complex with only additional learning, some appropriate forgetting is also necessary. It seems natural that many data are forgotten as the time elapses. We call the way of forgetting based only on the time elapse "passive forgetting". On the other hand, it is expected more effective to forget unnecessary data actively. We call this way of forgetting "unnecessary data" actively "active forgetting". In this research, several ways for active forgetting in machine leaning have been developed and applied to stock portfolio problems. As a result, it has been shown that active forgetting provides better results than mere additional learning or passive forgetting. It can be exp
ected that an effective decision support system for portfolio problems can be obtained by applying some of multi-objective programming techniques (e.g., Satisficing Trade-off Method developed by the author) to candidate stocks which are selected by machine learning with active forgetting.
In the first year of the research term, additional learning and passive forgetting in RBF networks was developed. Through numerical experiments, it was shown that this new technology works effectively in stock portfolio problems.
In the next year of the research term, rule extraction was tried by using the rough set theory. Although many machine learning techniques such as artificial neural networks can provide good results, they are not transparent (i.e., of black box). In many actual situations, people want to see how the prediction was made. To this end, extraction of explicit rules is needed. It was shown that the rough set theory can work effectively for this purpose.
In the last year of the research term, active forgetting was developed. Applying active forgetting in the potential method, remarkably beneficial results were obtained in stock portfolio problems. On the basis of the obtained results, a decision support system for stock portfolio is on trial to combine the above machine learning techniques and multi-objective programming techniques. Less