DRL rules from data (Drools-ML integration)
The idea is to combine domain expert's knowledge, that is available in a symbolic form (DRL rules), with rules that can be derived from data, using a machine learning algorithm, to build a more comprehensive Drools application.
The repo currently includes a very basic example that shows how a DRL can be validated against a given set of data, and how the emla libray can be used to discover additional rules for cases not covered by the DRL (if any). These rules can be combined with the initial DRL for the development of a more comprehensive Drools application. Incorporation of the new rules into the existing DRL is still under investigation.
The DroolsAgentApp example:
- The initial DRL contains the following rules:
rule AllowAdmin when
$a: Agent( role == "admin" )
then
$a.setGrantAccess( true );
allow.add( $a.getId() );
end
rule DenyGuest when
$a: Agent( role == "guest" )
then
$a.setGrantAccess( false );
deny.add( $a.getId() );
end
- A csv file contains 6 entries of agent request: 2 requests from
adminagents, 2 requests fromguestcases and 2 requests fromcontributorcases. - Data from the csv file are used to create
Agentobjects. - To validate the initial DRL:
for (Agent a : agentRequests){
kieSession.insert(a);
}
kieSession.fireAllRules();
if ( (allow.size()+deny.size()) < agentRequests.size()) {
// identify the cases not covered
// use emla to find additional rules for these cases
// revise the initial DRL
// repeat the validation process
}- The validation shows that the initial DRL cannot support the 2
contributorrequests. We can then useemlato derive a rule for these requests from the 2 dataset entries and add i.e. the following to the initial DRL:
rule AllowContributor when
$a: Agent( role == "contributor" )
then
$a.setGrantAccess( true );
allow.add( $a.getId() );
end
- Now the DRL contains three rules.
- If we repeat the validation process all
Agentrequests will be added in one of the lists,allowordeny.At this point we consider that the DRL fully covers the given dataset.