Improved the way models are created, by implementing a solution between TensorFlowJS and the user, giving interfaces to allow easier creation. Also, added support for Convolutional Neural Networks and for model loading from Keras or TensorFlow models.
- Added Neural Networks, a class which permits to create a neural network. This class corresponds to the pre-creation of the model, only giving each layer its configuration.
- Add layers to your neural network (dense, dropout, flatten)
- Added Convolutional Neural Network, an extension of the Neural Networks which permits to add convolutional layers (conv2d, maxpool2d).
- In the Convolutional Neural Network, the structure is managed automatically (conv layers => dense layers), you just have to set the content of each part.
- Convolutional Networks are not ready for use for now
This version corresponds to the minimal to have a working library.
- Create an academy, agents, teachers
- Associate teachers and agents
- Give an agent a neural network model
- Neural Network is managed by TensorFlowJS
- Implemented Q-Learning algorithm and here DQN
- Reward your agents
- Possibility to dynamically manage the learning sequence
- Parameters to create to learning sessions
- Possibility to visualize training sequence parameters by providing a debug output