zomis/machine-learning

Just me learning about Neural Networks and machine learning

machine-learning's Issues

Easier to classify programmers comments

Make it easier to classify stuff, right now with focus on classifying comments mentioning programmers.

While some work has been done already, also make it possible to continue from a specific chat message id, and remember which the most recent chat message id was.

Also make it possible to distribute the work among many people, so that it is faster to classify comments.

Learning curves

To detect possible bias and variance, make use of learning curves for the training set and cross-validation set.

Confusion Matrix

For multi-class classification (such as scanning a Minesweeper board), add use of Confusion Matrix

Implement regularization for Neural Networks

Add regularization for Neural Networks, in both cost function and gradient calculation of course.

Don't use both TrainingData and LearningData

Having both a TrainingData class and a LearningData class is confusing. Get rid of TrainingData, as that is only used for the gameai package.

Try other Machine Learning libraries

Look into Neuroph and Deep Learning 4j and measure the performances of those.

Use Matrixes for weights

Use Apache Commons RealMatrix and keep each layer's weights as a Matrix and use that Matrix to calculate activations and stuff.

Learn by opponent moves

Make the Game AI also learn from the moves that the opponent makes. This is especially important when playing against a better opponent (imagine an AI playing Game21 against an opponent that plays perfectly)

Use Apache Commons optimizers

Try using Apache Commons Optimizers instead of the current Backpropagation