I was having difficulty in training a policy for conversation of wildly different but deterministic flows, I think using a decison tree might help. Any idea how I can incorporate that?
I think you can use the Memoization policy to create deterministic flows, however it will not account for unlikely scenarios
tried that but memoization policy isnt able to learn exixsting flows, cant understand why.
Strange, set augmentation to 0, otherwise it randomly glues stories together
Thanks I will try that
didnt help, I have about 300~ stories
can you run with --debug to see what the policy predicts?
It just says
There is no memorised next action Predicted next action ‘action_listen’ with prob 0.00. Action ‘action_listen’ ended with events ‘’ topic: None
keep in mind, memoization looks at the entire history of the conversation in the tracker
if you want to ignore that, you have to create headless stories
_intent_hello - utter_hello
and set augmentation to 0
i like this explanation
class MemoizationPolicy(Policy): """The policy that remembers exact examples of `max_history` turns from training stories. Since `slots` that are set some time in the past are preserved in all future feature vectors until they are set to None, this policy implicitly remembers and most importantly recalls examples in the context of the current dialogue longer than `max_history`. This policy is not supposed to be the only policy in an ensemble, it is optimized for precision and not recall. It should get a 100% precision because it emits probabilities of 1.0 along it's predictions, which makes every mistake fatal as no other policy can overrule it. If it is needed to recall turns from training dialogues where some slots might not be set during prediction time, and there are training stories for this, use AugmentedMemoizationPolicy. """
I am somewhat confused what effect does max history have on memoization?
in order to predict the next action using this policy - it will compare the tracker with a max_history let’s say 1
* hello - utter_hello * how_are_you - utter_fine * what_can_you_do - help_with_mails
suppose this is your conversation with the user so far
when you use memoization with max_history 1, in order to predict what should happen next, it will consider what happened one history before and compare it with your training data so your training data must have
* what_can_you_do - help_with_mails * how_to_send_mail - send_by_button
then only send_by_button will be called as the next action for a particular intent
Tbh, i haven’t used Memoization just by itself but i use it to make sure that some really certain flows don’t break ( happy path ) and use Keras for the unlikely scenarios
But i will still answer your first question though,
You can override the model_architecture to implement your own policies for decision trees such as CART and pass your policy to the agent
from your policy import DecisionTree
from rasa_core.agent import Agent
agent = Agent(“domain.yml”, policies=[ DecisionTree(your_custom_features)])
what if I train the memoization policy for longer and predictable flows and separately train the keras policy for shorter/stochastic flows, and use them together. Could that work?
well technically an ensemble is using them together
For each prediction - the ensemble checks which gives a better prediction
You can pass both the policies to the agent, i am not sure how would you route your parse to two different models and check which is right