2015 Fiscal Year Research-status Report
Evolutionary Approaches to Learning Self-awareness for a Decentralized System
| Project/Area Number |
15K00343
|
|---|
| Research Institution | The University of Aizu |
|---|
Principal Investigator |
劉 勇 会津大学, コンピュータ理工学部, 准教授 (60325967)
|
|---|
| Project Period (FY) |
2015-04-01 – 2018-03-31
|
| Keywords | neural networks / awareness computing / ensemble learning |
|---|
| Outline of Annual Research Achievements |
This project is to establish a methodology for designing a decentralized learning system with a set of self-aware neural network subsystems. The hypothesis adopted in this project implies that systems being aware of their own state, behavior and performance can manage trade-offs between goals at run time. Such self-awareness enables systems to better meet their requirements in uncertain and dynamic environments.
Two levels of self-awareness have been created in the ensemble learning systems, including private self-awareness at the individual level and public self-awareness at the ensemble level. Private self-awareness at the individual level is generated by negative correlation learning with difference learning. The idea of difference learning is to let each individual in an ensemble learn to be different to the ensemble on some selected data points when the outputs of the ensemble are too close to the target values of these data. Public self-awareness at the ensemble level is created by the bounded negative correlation learning. There are two error bounds in the bounded negative correlation learning. One is the upper bound of error output which divides the training data into two groups based on the distances between the data and the formed decision boundary. The other is the lower bound of error rate which is set as a learning switch on the two groups of the training data.
Experimental results have been conducted to show how both difference learning and the bounded negative correlation learning could learn the given data faster and better.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The two key research problems raised in this projcet have been resolved by the bounded negative correlation learning and difference learning. The two problems are how each subsystem will obtain its internal knowledge and local environmental knowledge in a decentralized machine learning system, and how to achieve global objectives of high performance and fast learning through local actions and interactions.
The proposed negative correlation learning with difference learning and the bounded negative correlation learning have been implemented and tested on four real world data sets. The research results have been published and presented in the four international conferences.
|
| Strategy for Future Research Activity |
Negative correlation learning with self-awareness error functions will be developed. The self-awareness error functions could selectively increase the differences among the individuals in an ensemble so that the global objectives would be likely achieved.
In the original version of negative correlation learning, the differences are generated by minimizing the correlations between every pair of two individuals in the ensemble. It has been found that such differences would not be enough for preventing negative correlation learning from being overfitting if negative correlation learning would be conducted too long. The overfitting might appear at the last. In many real world applications, it needs that a learning system could continue to learn when the new data keep coming. Therefore, it would be desirable to design a learning system that could endure a long learning. One way to tolerate such a long learning is to make each individual learner have self-awareness so that not only would the error signals be learned, but also something such as the selective differences would be well enforced.
|
