Why Rasa/policies are inconsistent? I need some clarity on the below questions

Rasa Open Source
1

Hi! I’m still trying to figure out how does the Memoization and Ted policies work?.

sample story:

  • story: check steps:
    • intent: check_data
    • slot_was_set:
      • Role: null
    • action: action_auth
    • slot_was_set:
      • is_loggedin: True
    • action: action_chek_data
    • action: action_picker
    • intent: check_data
    • slot_was_set:
      • Role: seller
    • action: action_check_data2
    • intent: check_data
    • slot_was_set:
      • Role: seller
    • action: action_url
    • slot_was_set:
      • memberRole: null

Here are the questions…

  1. Ambiguities around memoization and why it does not work the way we expect it to: I want to understand why in cases that there is an exact match between user interaction and a user story, TED is the decision maker, but based on my understanding on Memoization policy Memoization should decide the next action in this case. Here till action_picker action memoization policy is predicting the next action, but after that point Ted policy is the decision maker, if after the interactions and user stories are matching, I’m still trying to understand why MemoizationPolicy is not applied here?

  2. In some cases ted policy is predicting the next action but with low confidence , what can be the possible reason for this behaviour?

  3. In some places flow is not working as expected(ted policy is failing to decide the next action) and Inconsistent behaviour of the trained model — One possible reason could be i did not had the much training examples to train the Ted policy, but can you please help me with how much and what of data we call as good data to train ted policy? and do we have any alternate solution?

  4. Even though the slot values are set as expected but still ted predicts the wrong actions. what can be the reason

2

There are any number of reasons this could be happening. The --debug option will show you the dialogue prediction.

However, in general, I would start with:

  • short rules instead of long stories (as you have above)
  • forms instead of featurized slots (as you have above)

The dialog prediction will be more reliable if you do these two things.