Add new search strategy: Large Neighborhood Search (LNS) #204
Conversation
… solution to returning model
src/CP/core/search/strategies.jl
Outdated
|
|
||
| """ | ||
| struct LNSearch <: SearchStrategy | ||
| implements the basic version of the Large Neighboorhood Search. |
There was a problem hiding this comment.
You can add a bit more doc about the Large Neighboorhood search or provide an external link
src/CP/core/search/lns.jl
Outdated
|
|
||
| ### Destroy and repair loop ### | ||
|
|
||
| while (isnothing(globalTimeLimit) || peektimer() < globalTimeLimit) && bestSolution[objective] > optimalScore |
There was a problem hiding this comment.
optimalScore can be deceiving. I 'd rather use optimalScoreLowerBound
src/CP/core/search/lns.jl
Outdated
| function destroy(model, solution, numberOfValuesToRemove, objective) | ||
|
|
||
| # Reset model | ||
| objectiveBound = solution[objective] - 1 | ||
| SeaPearl.reset_model!(model) | ||
|
|
||
| # Get variable fixed by current solution | ||
| vars = collect(values(model.variables)) | ||
| branchableVariablesId = collect(filter(e -> model.branchable[e], keys(model.branchable))) | ||
| varsToSet = sample(branchableVariablesId, count(values(model.branchable)) - numberOfValuesToRemove; replace=false) | ||
|
|
||
| # Fix some variables as in current solution | ||
| for var in varsToSet | ||
| variable = vars[findfirst(e -> e.id == var, vars)] | ||
| value = solution[var] | ||
| SeaPearl.assign!(variable, value) | ||
| end | ||
|
|
||
| # Pruning the objective domain to force the search for a better solution | ||
| objectiveVariable = vars[findfirst(e -> e.id == objective, vars)] | ||
| SeaPearl.removeAbove!(objectiveVariable.domain, objectiveBound) | ||
|
|
||
| return model | ||
| end |
There was a problem hiding this comment.
Some testsets on destroy function would be usefull
src/CP/core/search/lns.jl
Outdated
| # Get variable fixed by current solution | ||
| vars = collect(values(model.variables)) | ||
| branchableVariablesId = collect(filter(e -> model.branchable[e], keys(model.branchable))) | ||
| varsToSet = sample(branchableVariablesId, count(values(model.branchable)) - numberOfValuesToRemove; replace=false) |
There was a problem hiding this comment.
Done! I tried to add tests for the seedargument but I think there is no way to do it... But I have tested locally and it works properly!
src/CP/core/search/lns.jl
Outdated
| function repair(model, repairSearch, objective, variableHeuristic, valueSelection) | ||
| search!(model, repairSearch, variableHeuristic, valueSelection) | ||
| solutions = filter(e -> !isnothing(e), model.statistics.solutions) | ||
|
|
||
| if isempty(solutions) | ||
| toReturn = nothing | ||
| else | ||
| scores = map(solution -> solution[objective], solutions) | ||
| bestSolution = solutions[findfirst(score -> score == Base.minimum(scores), scores)] | ||
| toReturn = bestSolution | ||
| end | ||
| return toReturn | ||
| end |
There was a problem hiding this comment.
Some testsets on repair function would be usefull
test/CP/core/search/lns.jl
Outdated
|
|
||
| search = SeaPearl.LNSearch(limitValuesToRemove = 1) | ||
| status = SeaPearl.expandLns!(search, model, SeaPearl.MinDomainVariableSelection(), SeaPearl.BasicHeuristic()) | ||
| @test status === :NonOptimal |
There was a problem hiding this comment.
Can you recall why the result is not Optimal ?
There was a problem hiding this comment.
Because the first solution found (during initialization) will be {x => 2, y => 2} (because BasicHeuristic choose the biggest value in the domain). This solution is non optimal as x is the objective variable and his domain is {1, 2}. As we set limitValuesToRemove to 1, the destroy and repair loop will never reach any solution (because of the constraint Equal(x, y, trailer)). In that way, the solution returned will be {x => 2, y => 2} (≠ to optimal solution {x => 1, y => 1}).
There was a problem hiding this comment.
I will add some comments in test file
There was a problem hiding this comment.
Ok thanks, that's perfectly clear!
|
Last query : could you update some functions documentation to satisfy the standard function head where the arguments are individually listed and explained? |
|
Everything looks good for me ! You can squash & merge the PR ! |
Implementation of Large Neighborhood Search (LNS) to add a new search strategy to SeaPearl (existing strategies: DFS, IDS and RBS)