In this version, the internal data structure of snipe got some flexible extendability features. What's more, the training procedure resilient propagation was overhauled ( this includes an interface change). You can download snipe as usual by going to the snipe page.

• Dynamic synapse shadows have been added. The data structure was generalized (without slowing it down, no worries). But after all, what is a dynamic synapse shadow? It's pretty easy. In the documentation of Snipe, I mentioned a specially designed, pretty fast and flexible data structure for the storage of synapse data. A shadow now offers an additional double data cell for every single synapse. Arbitrarily many of such shadows can be added. This is very useful for e.g. training procedures like backpropagation, that need to store a learning rate or something els per synapse. Now that we know what a shadow is, the word “dynamic” remains to explain. If the network gets changed in structure (e.g. by removing a neuron), the shadows will be adapted automatically. At the moment, I implemented several operators to deal with synapse shadows, without really documenting them. However, they will be the fundament for a clean and slick inheritance policy which may be introduced in the next release.
• Resilient Backpropagation was overhauled. Note: This update causes a change in interface concerning the method com.dkriesel.snipe.core.NeuralNetwork.trainResilientBackpropagation. A big thank you for the overhaul goes to Martin Westhoven. He not only managed to fix a bug that might have prevented resilient propagation from learning (no other components were influenced), but also implemented a further Rprop-enhancement that you can activate if you like (have a look at the documentation). Furthermore, Rprop has been adapted to the synapse shadow structure.
• Several classes are serializable now.

Have a lot of fun!