Jump to navigation Jump to search
- It is a library I made for myself to use it.
- There aren't too many bots using NN's so I thought that people start to make more with a ready lib.
What can I do with it?
- You can create a Multi-Layer Perceptron.
- You can choose how many layers you'll use and how many nodes there will be.
- You can decide what learning rate you'll use.
- You can change the batch size.
- You can pre-process data and make it more useful for Neural Networks.
- You can choose different Activation Functions for each layer.
- There are activation functions coming with the package.
- Hyperbolic Tangent(Tanh)
- Sine(Perfect for overfitting).
- Also you can add your own activation functions.
- You can make a powerful GF gun.
- You can make a PIF gun by giving normal data + tick of the simulation.
- You can make a PM gun like in ScruchiPu.
- You can make a fire power choosing system.
- You can make a melee movement with regression. For example input as distance to objects in 360 different angle.
- You can create a MinimumRisk danger predictor.
- And lot more.
What to do next?
- Adding momentum.
- Adding recurrent neural networks.
- Adding auto-encoder which is already ready.
- Adding a Self-Organising Map.
- Adding softMax functions for deep learning maybe?
- Adding Deep Convolutional Inverse Graphics Deep Belief Network.(Just joking) =).
- Usage of Sigmoid or HyperbolicTangent is recommended. They have a protection against the vanishing gradient problem.
- Usage of Feature Splitter is recommended until I add Radial Basis Functions.
- Use SoftPlus(SmoothMax) if you need a result that is between 0-Positive Infinity.
Bots using Roboneural
Here is the link: https://www.dropbox.com/sh/8gwgik1gpvo4jzw/AAB6qYCt6fL-yTktysuzzRHBa?dl=1 If I was making a change when you were downloading, it may give an error.