private PartitionScore SplitOnAxis(ContinuousAxis axis, List <Episode> episodes, double[] totalWeights)
{
List <EvaluatedEpisode> line =
episodes
.AsParallel()
.Select(episode => EvaluateEpisode(episode, axis))
.OrderBy(x => x.Value)
.ToList();
int lastIndex = 0;
double[] tacticWeightsLeft = new double[NumTactics];
int bestSplitPosition = -1;
double bestEntropy = double.MaxValue;
for (int i = 1; i < line.Count; ++i)
{
bool sameAsPrevious = i > 0 && line[i - 1].Value == line[i].Value;
if (sameAsPrevious)
{
// Cannot create a split part way through the same number
continue;
}
while (lastIndex < i)
{
Episode episode = line[lastIndex++].Episode;
tacticWeightsLeft[(int)episode.Tactic] += episode.Weight;
}
double[] tacticWeightsRight = TacticEntropy.Subtract(totalWeights, tacticWeightsLeft);
double entropy = TacticEntropy.Entropy(tacticWeightsLeft, tacticWeightsRight);
if (entropy < bestEntropy)
{
bestEntropy = entropy;
bestSplitPosition = i;
}
}
if (bestSplitPosition == -1)
{
// Unable to find split - all values must be the same
return(null);
}
return(new PartitionScore {
Partitioner = new ContinuousPartitioner(axis, line[bestSplitPosition].Value),
LeftEpisodes = line.Take(bestSplitPosition).Select(x => x.Episode).ToList(),
RightEpisodes = line.Skip(bestSplitPosition).Select(x => x.Episode).ToList(),
Entropy = bestEntropy,
});
}
private DecisionNodeAccuracy OptimalLeaf(List <Episode> episodes, double[] tacticWeights)
{
TacticWeight bestTacticOverall =
tacticWeights
.Select((weight, i) => new TacticWeight {
Tactic = (Tactic)i, Weight = weight
})
.MaxByOrDefault(x => x.Weight);
if (!IsSpell(bestTacticOverall.Tactic))
{
return(new DecisionNodeAccuracy {
DecisionNode = new DecisionLeaf(bestTacticOverall.Tactic),
CorrectWeight = bestTacticOverall.Weight,
});
}
Dictionary <HeroType, Tactic> bestTacticPerHero = new Dictionary <HeroType, Tactic>();
double totalCorrectWeight = 0.0;
foreach (var group in episodes.GroupBy(ep => ep.Hero.HeroType))
{
TacticWeight bestTacticForHero =
TacticEntropy.TacticFrequency(group)
.Select((weight, i) => new TacticWeight {
Tactic = (Tactic)i, Weight = weight
})
.MaxByOrDefault(x => x.Weight);
bestTacticPerHero[group.Key] = bestTacticForHero.Tactic;
totalCorrectWeight += bestTacticForHero.Weight;
}
foreach (HeroType heroType in EnumUtils.GetEnumValues <HeroType>())
{
if (bestTacticPerHero.GetOrDefault(heroType) == bestTacticOverall.Tactic)
{
bestTacticPerHero.Remove(heroType);
}
}
return(new DecisionNodeAccuracy {
DecisionNode = new DecisionLeaf(bestTacticOverall.Tactic, bestTacticPerHero),
CorrectWeight = totalCorrectWeight,
});
}
private static AdditionalCategory EvaluateAdditionalCategory(Enum category, IEnumerable <Episode> additionalEpisodes, double[] totalWeights, double[] rightWeights)
{
double[] additionalWeights = TacticEntropy.TacticFrequency(additionalEpisodes);
double[] newRightWeights = new double[TacticEntropy.NumTactics];
double[] newLeftWeights = new double[TacticEntropy.NumTactics];
for (int i = 0; i < additionalWeights.Length; ++i)
{
newRightWeights[i] = rightWeights[i] + additionalWeights[i];
newLeftWeights[i] = totalWeights[i] - newRightWeights[i];
}
return(new AdditionalCategory {
Category = category,
AdditionalEpisodes = additionalEpisodes,
NewRightWeights = newRightWeights,
Entropy = TacticEntropy.Entropy(newLeftWeights, newRightWeights),
});
}
private async Task <DecisionNodeAccuracy> OptimalDecisionNode(List <Episode> episodes, double grandTotalWeight)
{
double[] tacticWeights = TacticEntropy.TacticFrequency(episodes);
double totalWeight = tacticWeights.Sum();
double leafEntropy = TacticEntropy.Entropy(tacticWeights);
DecisionNodeAccuracy bestLeaf = OptimalLeaf(episodes, tacticWeights);
if (bestLeaf.CorrectWeight + double.Epsilon >= totalWeight)
{
// We already know the perfect answer
return(bestLeaf);
}
double maximumCorrectWeight = totalWeight;
double maximumIncrease = maximumCorrectWeight - bestLeaf.CorrectWeight;
double requiredIncrease = RequiredIncreaseToSplit * grandTotalWeight;
if (maximumIncrease < requiredIncrease)
{
// Cannot make a big enough difference to the big picture
return(bestLeaf);
}
PartitionScore partitionScore =
_continuousLearner.Optimizers(episodes, tacticWeights)
.Concat(_categoricalLearner.Optimizers(episodes, tacticWeights))
.AsParallel()
.Select(partitionEvaluator => partitionEvaluator())
.Where(x => x != null && x.LeftEpisodes.Count > 0 && x.RightEpisodes.Count > 0)
.MinByOrDefault(x => x.Entropy);
if (partitionScore == null || partitionScore.Entropy >= leafEntropy)
{
if (partitionScore?.Entropy > leafEntropy + double.Epsilon)
{
Console.WriteLine("Entropy should never increase with a split");
}
// No split found
return(bestLeaf);
}
DecisionNodeAccuracy left = await OptimalDecisionNode(partitionScore.LeftEpisodes, grandTotalWeight);
DecisionNodeAccuracy right = await OptimalDecisionNode(partitionScore.RightEpisodes, grandTotalWeight);
DecisionNode splitter = new DecisionNode()
{
Partitioner = partitionScore.Partitioner,
Left = left.DecisionNode,
Right = right.DecisionNode,
};
double bestSplitCorrectWeight = left.CorrectWeight + right.CorrectWeight;
double correctWeightIncrease = bestSplitCorrectWeight - bestLeaf.CorrectWeight;
if (correctWeightIncrease < requiredIncrease)
{
return(bestLeaf);
}
else
{
return(new DecisionNodeAccuracy {
DecisionNode = splitter,
CorrectWeight = bestSplitCorrectWeight,
});
}
}