public virtual void EvalRoundRobin(
IEvolutionState state,
int[] origins,
int[] numinds,
IList <Individual> individuals,
int subpop,
IGroupedProblem prob)
{
if (state.EvalThreads == 1)
{
EvalRoundRobinPopChunk(
state, origins[0], numinds[0], 0, individuals, subpop, prob);
}
else
{
ParallelEvaluation <RoundRobinCompetitiveEvaluatorThread>(
state, origins, numinds, individuals, subpop, prob);
}
}
void ParallelEvaluation <TEvalThread>(
IEvolutionState state,
int[] origins,
int[] numinds,
IList <Individual> individuals,
int subpop,
IGroupedProblem prob)
where TEvalThread : CompetitiveEvaluatorThread, new()
{
// BRS: TPL DataFlow is cleaner and safer than using raw threads.
// Limit the concurrency in case the user has gone overboard!
var maxDegree = Math.Min(Environment.ProcessorCount, state.BreedThreads);
var options = new ExecutionDataflowBlockOptions {
MaxDegreeOfParallelism = maxDegree
};
Action <CompetitiveEvaluatorThread> act = t => t.Run();
var actionBlock = new ActionBlock <CompetitiveEvaluatorThread>(act, options);
for (var i = 0; i < state.EvalThreads; i++)
{
var runnable = new TEvalThread
{
ThreadNum = i,
State = state,
Subpop = subpop,
NumInds = numinds[i],
From = origins[i],
Problem = prob,
Evaluator = this,
Inds = individuals
};
actionBlock.Post(runnable);
}
actionBlock.Complete();
actionBlock.Completion.Wait();
}
public virtual void EvalNRandomTwoWayPopChunk(IEvolutionState state, int origin, int numinds, int threadnum,
IList <Individual> individuals, int subpop, IGroupedProblem prob)
{
// the number of games played for each player
var individualsOrdered = new EncapsulatedIndividual[individuals.Count];
var queue = new EncapsulatedIndividual[individuals.Count];
for (var i = 0; i < individuals.Count; i++)
{
individualsOrdered[i] = new EncapsulatedIndividual(individuals[i], 0);
}
var competition = new Individual[2];
var subpops = new[] { subpop, subpop };
var updates = new bool[2];
updates[0] = true;
var upperBound = origin + numinds;
for (var x = origin; x < upperBound; x++)
{
Array.Copy(individualsOrdered, 0, queue, 0, queue.Length);
competition[0] = queue[x].Ind;
// if the rest of individuals is not enough to fill
// all games remaining for the current individual
// (meaning that the current individual has left a
// lot of games to play versus players with index
// greater than his own), then it should play with
// all. In the end, we should check that he finished
// all the games he needs. If he did, everything is
// ok, otherwise he should play with some other players
// with index smaller than his own, but all these games
// will count only for his fitness evaluation, and
// not for the opponents' (unless allowOverEvaluations is set to true)
// if true, it means that he has to play against all opponents with greater index
if (individuals.Count - x - 1 <= GroupSize - queue[x].NumOpponentsMet)
{
for (var y = x + 1; y < queue.Length; y++)
{
competition[1] = queue[y].Ind;
updates[1] = queue[y].NumOpponentsMet < GroupSize || AllowOverEvaluation;
prob.Evaluate(state, competition, updates, false, subpops, 0);
queue[x].NumOpponentsMet++;
if (updates[1])
{
queue[y].NumOpponentsMet++;
}
}
}
// here he has to play against a selection of the opponents with greater index
else
{
// we can use the queue structure because we'll just rearrange the indexes
// but we should make sure we also rearrange the other vectors referring to the individuals
for (var y = 0; GroupSize > queue[x].NumOpponentsMet; y++)
{
// swap to the end and remove from list
var index = state.Random[0].NextInt(queue.Length - x - 1 - y) + x + 1;
competition[1] = queue[index].Ind;
updates[1] = queue[index].NumOpponentsMet < GroupSize || AllowOverEvaluation;
prob.Evaluate(state, competition, updates, false, subpops, 0);
queue[x].NumOpponentsMet++;
if (updates[1])
{
queue[index].NumOpponentsMet++;
}
// swap the players (such that a player will not be considered twice)
var temp = queue[index];
queue[index] = queue[queue.Length - y - 1];
queue[queue.Length - y - 1] = temp;
}
}
// if true, it means that the current player needs to play some games with other players with lower indexes.
// this is an unfortunate situation, since all those players have already had their groupSize games for the evaluation
if (queue[x].NumOpponentsMet < GroupSize)
{
for (var y = queue[x].NumOpponentsMet; y < GroupSize; y++)
{
// select a random opponent with smaller index (don't even care for duplicates)
int index;
if (x > 0)
{
// if x is 0, then there are no players with smaller index, therefore pick a random one
index = state.Random[0].NextInt(x);
}
else
{
index = state.Random[0].NextInt(queue.Length - 1) + 1;
}
// use the opponent for the evaluation
competition[1] = queue[index].Ind;
updates[1] = (queue[index].NumOpponentsMet < GroupSize) || AllowOverEvaluation;
prob.Evaluate(state, competition, updates, false, subpops, 0);
queue[x].NumOpponentsMet++;
if (updates[1])
{
queue[index].NumOpponentsMet++;
}
}
}
}
}
public virtual void EvalNRandomOneWayPopChunk(IEvolutionState state, int origin, int numinds, int threadnum, IList <Individual> individuals, int subpop, IGroupedProblem prob)
{
var queue = individuals.ToArray();
int len = queue.Length;
var competition = new Individual[2];
var subpops = new[] { subpop, subpop };
var updates = new bool[2];
updates[0] = true;
updates[1] = false;
var upperBound = origin + numinds;
for (var x = origin; x < upperBound; x++)
{
competition[0] = individuals[x];
// fill up our tournament
for (var y = 0; y < GroupSize;)
{
// swap to end and remove
var index = state.Random[0].NextInt(len - y);
competition[1] = queue[index];
queue[index] = queue[len - y - 1];
queue[len - y - 1] = competition[1];
// if the opponent is not the actual individual, we can
// have a competition
if (!competition[1].Equals(individuals[x]))
{
prob.Evaluate(state, competition, updates, false, subpops, 0);
y++;
}
}
}
}
/// <summary>
/// A private helper function for evalutatePopulation which evaluates a chunk
/// of individuals in a subpop for a given thread.
///
/// Although this method is declared public (for the benefit of a private
/// helper class in this file), you should not call it.
/// </summary>
public virtual void EvalRoundRobinPopChunk(IEvolutionState state, int origin, int numinds, int threadnum, IList <Individual> individuals, int subpop, IGroupedProblem prob)
{
var competition = new Individual[2];
var subpops = new[] { subpop, subpop };
var updates = new bool[2];
updates[0] = updates[1] = true;
var upperBound = origin + numinds;
// evaluate chunk of population against entire population
// since an individual x will be evaluated against all
// other individuals <x in other threads, only evaluate it against
// individuals >x in this thread.
for (var x = origin; x < upperBound; x++)
{
for (var y = x + 1; y < individuals.Count; y++)
{
competition[0] = individuals[x];
competition[1] = individuals[y];
prob.Evaluate(state, competition, updates, false, subpops, 0);
}
}
}
public virtual void EvalSingleElimination(IEvolutionState state, IList <Individual> individuals, int subpop, IGroupedProblem prob)
{
// for a single-elimination tournament, the subpop[0] size must be 2^n for
// some value n. We don't check that here! Check it in Setup.
// create the tournament array
var tourn = individuals.ToArray();
var len = tourn.Length;
var competition = new Individual[2];
var subpops = new[] { subpop, subpop };
var updates = new bool[2];
updates[0] = updates[1] = true;
// the "top half" of our array will be losers.
// the bottom half will be winners. Then we cut our array in half and repeat.
while (len > 1)
{
for (var x = 0; x < len / 2; x++)
{
competition[0] = tourn[x];
competition[1] = tourn[len - x - 1];
prob.Evaluate(state, competition, updates, true, subpops, 0);
}
for (var x = 0; x < len / 2; x++)
{
// if the second individual is better, or coin flip if equal, than we switch them around
if (tourn[len - x - 1].Fitness.BetterThan(tourn[x].Fitness) ||
tourn[len - x - 1].Fitness.EquivalentTo(tourn[x].Fitness) && state.Random[0].NextBoolean())
{
Individual temp = tourn[x];
tourn[x] = tourn[len - x - 1];
tourn[len - x - 1] = temp;
}
}
// last part of the tournament: deal with odd values of len!
if (len % 2 != 0)
{
len = 1 + len / 2;
}
else
{
len /= 2;
}
}
}
public virtual void AfterCoevolutionaryEvaluation(IEvolutionState state, Population pop, IGroupedProblem prob)
{
if (NumElite > 0)
{
for (var i = 0; i < state.Population.Subpops.Count; i++)
{
if (ShouldEvaluateSubpop(state, i, 0)) // only load elites for subpopulations which are actually changing
{
LoadElites(state, i);
}
}
}
// copy over the previous population
if (NumPrev > 0)
{
_previousPopulation = (Population)(state.Population.EmptyClone());
for (var i = 0; i < _previousPopulation.Subpops.Count; i++)
{
for (var j = 0; j < _previousPopulation.Subpops[i].Individuals.Count; j++)
{
_previousPopulation.Subpops[i].Individuals[j] = (Individual)(state.Population.Subpops[i].Individuals[j].Clone());
}
}
}
}
public virtual void PerformCoevolutionaryEvaluation(IEvolutionState state, Population pop, IGroupedProblem prob)
{
var evaluations = 0;
_inds = new Individual[pop.Subpops.Count];
_updates = new bool[pop.Subpops.Count];
// we start by warming up the selection methods
if (NumCurrent > 0)
{
for (var i = 0; i < _selectionMethodCurrent.Length; i++)
{
_selectionMethodCurrent[i].PrepareToProduce(state, i, 0);
}
}
if (NumPrev > 0)
{
for (var i = 0; i < _selectionMethodPrev.Length; i++)
{
// do a hack here
var currentPopulation = state.Population;
state.Population = _previousPopulation;
_selectionMethodPrev[i].PrepareToProduce(state, i, 0);
state.Population = currentPopulation;
}
}
// build subpopulation array to pass in each time
var subpops = new int[state.Population.Subpops.Count];
for (var j = 0; j < subpops.Length; j++)
{
subpops[j] = j;
}
// handle shuffled always
if (NumShuffled > 0)
{
int[] /*numShuffled*/ [] /*subpop*/ [] /*shuffledIndividualIndexes*/ ordering = null;
// build shuffled orderings
ordering = TensorFactory.Create <Int32>(NumShuffled, state.Population.Subpops.Count,
state.Population.Subpops[0].Individuals.Count);
for (var c = 0; c < NumShuffled; c++)
{
for (var m = 0; m < state.Population.Subpops.Count; m++)
{
for (var i = 0; i < state.Population.Subpops[0].Individuals.Count; i++)
{
ordering[c][m][i] = i;
}
if (m != 0)
{
Shuffle(state, ordering[c][m]);
}
}
}
// for each individual
for (var i = 0; i < state.Population.Subpops[0].Individuals.Count; i++)
{
for (var k = 0; k < NumShuffled; k++)
{
for (var ind = 0; ind < _inds.Length; ind++)
{
_inds[ind] = state.Population.Subpops[ind].Individuals[ordering[k][ind][i]];
_updates[ind] = true;
}
prob.Evaluate(state, _inds, _updates, false, subpops, 0);
evaluations++;
}
}
}
// for each subpopulation
for (var j = 0; j < state.Population.Subpops.Count; j++)
{
// now do elites and randoms
if (!ShouldEvaluateSubpop(state, j, 0))
{
continue; // don't evaluate this subpopulation
}
// for each individual
for (var i = 0; i < state.Population.Subpops[j].Individuals.Count; i++)
{
var individual = state.Population.Subpops[j].Individuals[i];
// Test against all the elites
for (var k = 0; k < _eliteIndividuals[j].Length; k++)
{
for (var ind = 0; ind < _inds.Length; ind++)
{
if (ind == j)
{
_inds[ind] = individual;
_updates[ind] = true;
}
else
{
_inds[ind] = _eliteIndividuals[ind][k];
_updates[ind] = false;
}
}
prob.Evaluate(state, _inds, _updates, false, subpops, 0);
evaluations++;
}
// test against random selected individuals of the current population
for (var k = 0; k < NumCurrent; k++)
{
for (var ind = 0; ind < _inds.Length; ind++)
{
if (ind == j)
{
_inds[ind] = individual;
_updates[ind] = true;
}
else
{
_inds[ind] = ProduceCurrent(ind, state, 0);
_updates[ind] = true;
}
}
prob.Evaluate(state, _inds, _updates, false, subpops, 0);
evaluations++;
}
// Test against random selected individuals of previous population
for (int k = 0; k < NumPrev; k++)
{
for (int ind = 0; ind < _inds.Length; ind++)
{
if (ind == j)
{
_inds[ind] = individual;
_updates[ind] = true;
}
else
{
_inds[ind] = ProducePrevious(ind, state, 0);
_updates[ind] = false;
}
}
prob.Evaluate(state, _inds, _updates, false, subpops, 0);
evaluations++;
}
}
}
// now shut down the selection methods
if (NumCurrent > 0)
{
for (var i = 0; i < _selectionMethodCurrent.Length; i++)
{
_selectionMethodCurrent[i].FinishProducing(state, i, 0);
}
}
if (NumPrev > 0)
{
for (var i = 0; i < _selectionMethodPrev.Length; i++)
{
// do a hack here
var currentPopulation = state.Population;
state.Population = _previousPopulation;
_selectionMethodPrev[i].FinishProducing(state, i, 0);
state.Population = currentPopulation;
}
}
state.Output.Message("Evaluations: " + evaluations);
}
public virtual void BeforeCoevolutionaryEvaluation(IEvolutionState state, Population pop, IGroupedProblem prob)
{
if (state.Generation == 0)
{
// create arrays for the elite individuals in the population at the previous generation.
// deep clone the elite individuals as random individuals (in the initial generation, nobody has been evaluated yet).
// deal with the elites
_eliteIndividuals = TensorFactory.Create <Individual>(state.Population.Subpops.Count, NumElite);
// copy the first individuals in each subpop (they are already randomly generated)
for (var i = 0; i < _eliteIndividuals.Length; i++)
{
if (NumElite > state.Population.Subpops[i].Individuals.Count)
{
state.Output.Fatal("Number of elite partners is greater than the size of the subpop.");
}
for (var j = 0; j < NumElite; j++)
{
_eliteIndividuals[i][j] = (Individual)(state.Population.Subpops[i].Individuals[j].Clone()); // just take the first N individuals of each subpopulation
}
}
// test for shuffled
if (NumShuffled > 0)
{
var size = state.Population.Subpops[0].Individuals.Count;
for (var i = 0; i < state.Population.Subpops.Count; i++)
{
if (state.Population.Subpops[i].Individuals.Count != size)
{
state.Output.Fatal("Shuffling was requested in MultiPopCoevolutionaryEvaluator, but the subpopulation sizes are not the same. " +
"Specifically, subpopulation 0 has size " + size + " but subpopulation " + i + " has size " + state.Population.Subpops[i].Individuals.Count);
}
}
}
}
}