A simple neural network library for training Deep Q Networks in different environments.
- 2D perlin noise
- Maze (generated using recursive backtracking)
- Pseudorandom noise
- DQN
- Double DQN
- Replay Buffer
- Prioritized Experience Replay
- Adam
public class Main {
public static void main(String[] args) {
Environment.setStateType(Environment.StateType.PositionVectorOnly);
Environment.setDimensions(10, 10);
Environment.setActionSpace(4);
DDQNAgentTrainer trainer;
try {
trainer = new DDQNAgentTrainer(Set.of(EmptyGridEnvironment.class, RandomGridEnvironment.class, PerlinGridEnvironment.class, MazeGridEnvironment.class));
} catch (InvalidTypeException e) {
e.printStackTrace();
return;
}
List<Layer> layers = new ArrayList<>();
LeakyReLU leakyRelu = new LeakyReLU(0.1f);
float lambda = 0.0001f;
// StateSpace is 104, ActionSpace is 5
layers.add(new MLPLayer(Environment.getStateSpace(), 64, leakyRelu, 0, lambda));
layers.add(new MLPLayer(64, 64, leakyRelu, 0, lambda));
layers.add(new MLPLayer(64, Environment.getActionSpace(), new Linear(), 0, lambda));
DDQNAgent ddqnAgent = new DDQNAgent(
Environment.getActionSpace(), // action space
layers, // layers
1, // initial epsilon
0.9999, // epsilon decay
0.01, // epsilon min
0.999, // gamma
0.0001, // learning rate
0.99995, // learning rate decay
0.000001f, // learning rate minimum
0.005 // tau
);
trainer.trainAgent(
ddqnAgent, // agent
600000, // num episodes
500, // save period
1, // visualiser update period
"plot", "ease", "axis_ticks", "show_path", "verbose" // varargs
);
}
}This list is incomplete, but I will try and ensure I add all the sources I used eventually
- https://arxiv.org/pdf/1511.08458
- http://arxiv.org/pdf/1511.05952v4
- https://arxiv.org/pdf/1412.6980
- http://arxiv.org/pdf/1502.03167
- https://proceedings.neurips.cc/paper/2020/file/32fcc8cfe1fa4c77b5c58dafd36d1a98-Paper.pdf
- https://doi.org/10.1109/ACCESS.2019.2941229
- http://www.arxiv.org/pdf/1509.06461
- https://arxiv.org/pdf/1312.5602