I don’t know how to do this but I’m interested too, if anyone can answer it.
About confidence in different policies: FallbackPolicy is set to fire whenever core_threshold is below the minimum in your settings. So if it’s fired, it means the other dialogue policies weren’t able to predict intent higher than the minimum.
We are currently facing this issue in our project now and I guess there are many things we should explore to improve it. The ones we are trying to tackle are: better policy setup, testing with different pipelines, more stories and adding entities (we’re currently trying to predict without entities yet, only intents). Let me know what you think you’re going to do too 
Also, here in documentation you can see how to change this core_threshold.