CONCLUSION

The central idea of this paper is to construct an adaptive system which learns to predict the effects of further learning. This is done by training an adaptive sub-module to predict (the expectation of the sum of) future error changes caused by a particular learning algorithm. Here one adaptive module learns to make estimates about the effects of the learning procedure of another adaptive module. In other words, there is a module which learns to make a statement about learning itself. This is related to the concept of `meta-learning': In a very limited sense the system learns how to learn.

The method represents a general strategy for learning to select training examples such that the expected performance improvement is maximized. Therefore the usefulness of the approach is not limited to model-building control systems. The principles above are general enough to be of interest whenever the task is to select appropriate training examples for any kind of learning system.

Back to Active Learning - Exploration - Curiosity page
Back to Reinforcement Learning page