Discussing the article: "MQL5 Wizard Techniques you should know (Part 60): Inference Learning (Wasserstein-VAE) with Moving Average and Stochastic Oscillator Patterns"

[MetaQuotes]

Check out the new article: MQL5 Wizard Techniques you should know (Part 60): Inference Learning (Wasserstein-VAE) with Moving Average and Stochastic Oscillator Patterns.

We wrap our look into the complementary pairing of the MA & Stochastic oscillator by examining what role inference-learning can play in a post supervised-learning & reinforcement-learning situation. There are clearly a multitude of ways one can choose to go about inference learning in this case, our approach, however, is to use variational auto encoders. We explore this in python before exporting our trained model by ONNX for use in a wizard assembled Expert Advisor in MetaTrader.

The use of the name ‘states’ to forecast price changes is fortuitous because from supervised-learning we move to reinforcement-learning. As it is established within reinforcement-learning, states are a key starting point to the training process, that much resembles the diagram below.

There are a number of variations in reinforcement-learning depending on the algorithm used but for most, in principle, they use two networks. The first being a policy that is shown as the upper of the two networks in the diagram above and the other being the value network which is represented as the lower.

Reinforcement can be a sole training method for a model or system, but we argued in the last article that it could be used more on live deployed models. When this is done, the exploration/ exploitation balance would be more relevant in ensuring an already trained model adapts to changing market environments. But even more than that, we saw how decisions to go long or short can be processed further in selecting the kind of action necessary for a forecast state.

Author: Stephen Njuki