Could you please let me know how are you identifying the features you're passing to keras policy. I could see story graphs are being created during agent.load_data.
Could you please share live example so that I can tune the parameters and hyperparameters to get the best out of keras lstm model.
Rasa Core version: 0.11.12
Python version: 3.5
Operating system (windows, osx, ...):windows 10
The selected feature depends on the used policy and its configuration. You can specify your configuration in a policy configuration file. If you use the "embedding policy", you can also define the layers etc. of the used LSTM in this configuration file.
The features are identified from
Have a look at the documentation on featurization for more details since this highly depends on the used policy configuration (you can select different featurizers).
With rasa_core version 0.12 you can only compare the accuracy of different policies with the command python -m rasa_core.train compare. This is probably helpful if you want to finetune that.