Implement Delayed Learning in adaptive model
Hi,
I need to implement delayed learning in adaptive model.
I have read about the various learning options and how PEGA implements delayed learning by utilising Cassandra. I have built a request and a response data flow for feeding the models.
So as I understand, during making a request, Cassandra will store the interactionID(which should not go to adaptive Model data set) and during response this interaction should be mapped with the incoming request and then only the response should be fed to adaptive.
Here the question is , during the time of making decision, when predictor inputs are recorded, is there a way to run an activity/data flow which will feed the adaptive model in case no response was received within the configured time (TimeToLive)? So if there was no response within that time, how does adaptive (Adaptive data set) get updated after the TimeToLive expires?
Thanks in advance.
I second Otto's comments. You can have a look at the following to know what the feature is. The delayed learning portion is automatically handled if you are using this feature.
Configuring outcome optimization strategies -
https://community.pega.com/knowledgebase/articles/configuring-outcome-optimization-strategies
Pega 8.1 Update: What's New in Outcome Optimization -
https://community.pega.com/video-library/pega-81-update-whats-new-outcome-optimization
Hi All,
Thank you for you answers.
I have one more question.
How is Cassandra data removed (or marked expired) after the Time To Live expires?
Is there any agent which is updating the pxDecisionresults (& Event Store) after the TTL expires?
Thank you in advance!
The Cassandra compaction process will remove the expired data,
Ivar
Good questions.
The most common pattern to handle implicit negatives is to post a negative at decision time. Yes this means double counting of the positives but when you use it for ranking, that does not influence the ranking.
The "royal" way is to introduce a time-out for both positives and negatives, so queue them up, then only after that time-out feed them to adaptive. This pattern has been productized in the so-called "outcome optimizer" which is part of NBA designer. Under the covers, this uses an ESM window to keep the responses and only send the last one (i.e. the positive one if there is one, otherwise the earlier posted negative) to ADM.