Description of problem:
When I register my own step listener and I ask AbstractStepScope if score was improved NPE is thrown during Construction Heuristic phase. The problem is that bestScoreImproved attribute is initialized after first LocalSearch step. But technically it should be false from the beginning because ConstructionHeuristic always lowers score.
Extends SolverPhaseLifecycleListener and in stepEnded method use for example really simple IF: if(stepScope.getBestScoreImproved().booleanValue())
NPE is thrown
result of calling .getBestScoreImproved().booleanValue() should be false
This is a known issue (which might not be solvable even). Note that LocalSearchPhaseLifecycleListener (unlike SolverEventListener) isn't part of the public API because they are deep inside the optaplanner-core. Despite that, the issue is valid of course.
In every step, the Construction Heuristics initialize 1 entity. Because this causes the uninitializedEntityCount to decrease, this actually means that the best score improved(!). Think about it, a solution with score -7 and 10 unassigned nurses is better than a solution with score 0 and 20 unassigned nurses.
However, this would imply that best solution changed, so the working solution needs to be cloned and an best solution event needs to be fired. That cloning is a huge performance killer. So CH's don't do that, they delay that until the last step...
Meanwhile, the bestSolutionImproved and other related field stay null to fail fast if any code is using them (which is better than holding the wrong non-null value...)
Going to reject this issue because reasons stated above and more importantly because the code does:
Downcasting to DefaultSolver, to be able to mess with the internal classes is not supported.
Meanwhile, Jiri's PR will improve the javadoc for those that do mess with the internal classes. Also, we 're investigating why there was the need to mess with the internal classes in the first place, and if we remove that need by supporting a new requirement on the public api.