toucheatout 2006-03-15 19:33 Machine Learning
Machine learning is a field of AI where the aim is to build a model of a problem until an acceptable solution is found. It is usually constructed on working with inputs and desired output from a test set (supervised learning, as decision trees, neural networks, rule-based systems, bayes networks, genetic algorithms) or input only (unsupervised learning, as clustering).
Once the model built, it can be used with new input data with unknown output to give predictions.
Most exciting research come from model agregation (through voting for instance, or through several levels hierarchical architecture, mix of both).
Common techniques
Dataset management
When working with a dataset of previous data to build a model, a portion of it is usually taken apart from learning to validate the outcome of the learning algorithm. To this end, instead of taking a fixed set of samples to validate learning, several advanced methods are now standard.
- N-fold cross-validation
- Bagging
- Boosting
Common problems with learning algorithms
- overfitting
- This is common to every supervised learning algorithm. Overfitting occurs when training on the data set has been excessive. The performance on the specific training samples is very good, but when presented to new data, the system will perform poorly.
- Bias
- Timed-based rewards
- When the correct output is determined not by a single-step decision, but arises after several non-rewarding steps (as for (classical) example of finding food in a maze: only the last step is rewarded, even though all the previous steps are necessary and are the major part determining the performance - e.g. minimum number of moves). The question is then how to learn the non-directly rewarding steps that allow for closing in the solution most efficiently.
Publicly available papers
- Christopher J.C. Burges
- A Tutorial on SVM.
- Altun, Tsochantaridis & Hofmann
- Combination of HMM and SVM.