Support learning from multiple episodes

[bvdmitri]

The current API infrastructure does not support learning from multiple episodes

[bvdmitri]

This is on hold

[bvdmitri]

Currently we store parameters in the model structure, and learning on different episodes will simply overwrite the previous one but actually also use it as the prior. It is problematic because learning from the same episode twice will effectively corrupt the learned parameters.

What would happen is that it will learn again by using the posteriors from the first learning as the prior. What should happen is the learned parameters should not change if learning from the same episode twice.