Code for the self reinforcement learning competition at www.blackboxchallenge.com
This is a self reinforcement learning submission using a neural network. Most of the tricks from DeepMinds Atari paper are included in some form. If there is structure in the data it should find it eventually :-) One trick I'm using (without success it seems) is to prime the network with both random exploration as well as optimal actions using checkpoints.
It uses a 100 neuron neural network:
- Has one value for a number of learning epochs with a high (but decreasing) exploration rate
- One value for a set number of learning epochs with a fixed exploration rate
- Stores neural network architecture and updates weights each epoch, so learning/training can be continued at will
- Repeats same action a specified number of times
- Updates the network at a specified interval