Based on Reinforcement Learning: Q-Learning with Clockwork RNN we develop the Futures-Trading-Robot. There are four main features below,
- Clockwork RNN is modified to two hidden layers.
- We adopt Double Dueling-DQN instead of DQN to improve the robustness of trading performance.
- For exploration and exploitation on model stability as training, boltzmann_policy is better than greedy_policy.
- Considering time series, data dependency, we replace experience replay with data incremental method.
- A clockwork RNN, https://arxiv.org/abs/1402.3511
- Dueling Network Architectures for Deep Reinforcement Learning, http://proceedings.mlr.press/v48/wangf16.pdf
- Exploration in DeepReinforcement Learning, https://www.ias.informatik.tu-darmstadt.de/uploads/Theses/Abschlussarbeiten/markus_semmler_bsc.pdf