Reinforcement Learning with TensorFlow

Further improvements

There are further improvements that can be made to the previous framework, and also better approaches to creating end to end financial portfolio managing agents using deep reinforcement learning. They are as follows:

• Current framework assumptions, which are zero slippage and zero market impact. Thus, considering market impact and slippage will provide real-world trading samples, which will improve the training dataset.
• Use of an actor-critic type of framework will help more in long-term market reactions.
• Preferring LSTMs and GRUs over basic RNNs overcomes the issue of the vanishing gradient problem.