/// <inheritdoc />
public void DetermineSampleSizes(SearchContext <List <SabberStoneAction>, SabberStoneState, SabberStoneAction, object, SabberStoneAction> context, out int generationSamples, out int evaluationSamples)
{
long estimatedSamples;
switch (BudgetType)
{
case BudgetType.Iterations:
// If we're running on an iterations budget, we know exactly how many we can spend.
estimatedSamples = BudgetAllowance;
break;
case BudgetType.Time:
// If we're running on a time budget, we can check what the previous search was able to run.
if (PreviousSearchTime > 0 && PreviousSearchIterations > 0)
{
var sampleDuration = PreviousSearchTime / (PreviousSearchIterations * 1.0);
// We can now estimate how many samples go into the budget we are allowed to spend
estimatedSamples = (long)(BudgetAllowance / (sampleDuration * AverageSampleTimeBudgetEstimationStrategy.TIME_BUDGET_SAFETY_MARGIN));
}
else
{
// If we don't have any statistics on the previous search (maybe cause this is going to be the first)
// use an estimation strategy based on a couple playouts
estimatedSamples = AverageSampleTimeBudgetEstimationStrategy.EstimateBudgetWithTestPlayouts(context, Playout, BudgetAllowance);
}
break;
default:
throw new InvalidEnumArgumentException($"BudgetType `{BudgetType}' is not supported.");
}
generationSamples = (int)(estimatedSamples * GenerationBudgetPercentage);
evaluationSamples = (int)(estimatedSamples * (1 - GenerationBudgetPercentage) * Constants.DEFAULT_LSI_EVALUATION_SAMPLES_ADJUSTMENT_FACTOR);
}
/// <summary>
/// Constructs a new instance of LSIBot with default strategies.
/// </summary>
/// <param name="allowPerfectInformation">[Optional] Whether or not this bot is allowed perfect information about the game state (i.e. no obfuscation and therefore no determinisation). Default value is false.</param>
/// <param name="ensembleSize">[Optional] The size of the ensemble to use. Default value is 1.</param>
/// <param name="playoutBotType">[Optional] The type of playout bot to be used during playouts. Default value is <see cref="PlayoutBotType.MAST"/>.</param>
/// <param name="mastSelectionType">[Optional] The type of selection strategy used by the M.A.S.T. playout. Default value is <see cref="MASTPlayoutBot.SelectionType.EGreedy"/>.</param>
/// <param name="playoutTurnCutoff">[Optional] The amount of turns after which to stop a simulation. Default value is <see cref="Constants.DEFAULT_PLAYOUT_TURN_CUTOFF"/>.</param>
/// <param name="budgetType">[Optional] The type of budget that this bot will use. Default value is <see cref="BudgetType.Iterations"/>.</param>
/// <param name="samples">[Optional] The budget for the amount of iterations LSI can use. Default value is <see cref="Constants.DEFAULT_COMPUTATION_ITERATION_BUDGET"/>.</param>
/// <param name="time">[Optional] The budget for the amount of milliseconds LSI can spend on searching. Default value is <see cref="Constants.DEFAULT_COMPUTATION_TIME_BUDGET"/>.</param>
/// <param name="generationBudgetPercentage">[Optional] The percentage of the budget that should be spent during the generation phase. Default value is <see cref="Constants.DEFAULT_LSI_BUDGET_GENERATION_PERCENTAGE"/>.</param>
/// <param name="budgetEstimationType">[Optional] The type of strategy used to estimate the budget for LSI. Default value is <see cref="BudgetEstimationType.AverageSampleTime"/>.</param>
/// <param name="useHeuristicEvaluation">[Optional] Whether or not to use the HeuristicBot's evaluation function. Default value is false.</param>
/// <param name="debugInfoToConsole">[Optional] Whether or not to write debug information to the console. Default value is false.</param>
public LSIBot(bool allowPerfectInformation = false,
int ensembleSize = 1,
PlayoutBotType playoutBotType = PlayoutBotType.MAST,
MASTPlayoutBot.SelectionType mastSelectionType = MASTPlayoutBot.SelectionType.EGreedy,
int playoutTurnCutoff = Constants.DEFAULT_PLAYOUT_TURN_CUTOFF,
BudgetType budgetType = BudgetType.Iterations,
int samples = Constants.DEFAULT_COMPUTATION_ITERATION_BUDGET,
long time = Constants.DEFAULT_COMPUTATION_TIME_BUDGET,
double generationBudgetPercentage = Constants.DEFAULT_LSI_BUDGET_GENERATION_PERCENTAGE,
BudgetEstimationType budgetEstimationType = BudgetEstimationType.AverageSampleTime,
bool useHeuristicEvaluation = false,
bool debugInfoToConsole = false)
{
PerfectInformation = allowPerfectInformation;
EnsembleSize = ensembleSize;
PlayoutBotType = playoutBotType;
MASTSelectionType = mastSelectionType;
PlayoutTurnCutoff = playoutTurnCutoff;
BudgetType = budgetType;
Samples = samples;
Time = time;
GenerationBudgetPercentage = generationBudgetPercentage;
BudgetEstimation = budgetEstimationType;
_debug = debugInfoToConsole;
// Create the ensemble search
Ensemble = new EnsembleStrategySabberStone(enableStateObfuscation: true, enablePerfectInformation: PerfectInformation);
// Adjust sample sizes for use in the Ensemble
long budgetAllowance;
switch (BudgetType)
{
case BudgetType.Iterations:
Samples = EnsembleSize > 0 ? Samples / EnsembleSize : Samples; // Note: Integer division by design.
budgetAllowance = Samples;
break;
case BudgetType.Time:
Time = EnsembleSize > 0 ? Time / EnsembleSize : Time; // Note: Integer division by design.
budgetAllowance = Time;
break;
default:
throw new InvalidEnumArgumentException($"BudgetType `{BudgetType}' is not supported.");
}
// Simulation will be handled by the Playout.
var sabberStoneStateEvaluation = new EvaluationStrategyHearthStone(useHeuristicEvaluation);
Playout = new PlayoutStrategySabberStone();
// Set the playout bots
switch (PlayoutBotType)
{
case PlayoutBotType.Random:
MyPlayoutBot = new RandomBot(filterDuplicatePositionTasks: true);
OpponentPlayoutBot = new RandomBot(filterDuplicatePositionTasks: true);
break;
case PlayoutBotType.Heuristic:
MyPlayoutBot = new HeuristicBot();
OpponentPlayoutBot = new HeuristicBot();
break;
case PlayoutBotType.MAST:
MyPlayoutBot = new MASTPlayoutBot(MASTSelectionType, sabberStoneStateEvaluation);
OpponentPlayoutBot = new MASTPlayoutBot(MASTSelectionType, sabberStoneStateEvaluation);
break;
default:
throw new InvalidEnumArgumentException($"PlayoutBotType `{PlayoutBotType}' is not supported.");
}
// Set the budget estimation strategy.
switch (BudgetEstimation)
{
case BudgetEstimationType.AverageSampleTime:
BudgetEstimationStrategy = new AverageSampleTimeBudgetEstimationStrategy(BudgetType, budgetAllowance, GenerationBudgetPercentage, Playout);
break;
case BudgetEstimationType.PreviousSearchAverage:
BudgetEstimationStrategy = new PreviousSearchAverageBudgetEstimationStrategy(BudgetType, budgetAllowance, GenerationBudgetPercentage, Playout);
break;
default:
throw new InvalidEnumArgumentException($"BudgetEstimationType `{budgetEstimationType}' is not supported.");
}
// LSI will need a goal-strategy to determine when a simulation is done
Goal = new GoalStrategyTurnCutoff(PlayoutTurnCutoff);
// LSI will need an evaluation-strategy to evaluate the strength of samples
Evaluation = new EvaluationStrategyHearthStone();
// Application will be handled by the GameLogic
// Hierarchical Expansion is set to TRUE because of incremental task validation during action sampling.
GameLogic = new SabberStoneGameLogic(Goal, hierarchicalExpansion: true);
// The side information strategy needs access to several of these.
SideInformationStrategy = new SabberStoneSideInformationStrategy(Playout, Evaluation, GameLogic);
// The sampling strategy used to sample actions during the generation phase.
SamplingStrategy = new SabberStoneLSISamplingStrategy(GameLogic);
}
