/// <summary>
/// Selects an action using the e-greedy algorithm.
/// </summary>
/// <param name="state">The game state.</param>
/// <returns><see cref="SabberStoneAction"/>.</returns>
private SabberStoneAction SelectEGreedy(SabberStoneState state)
{
// Determine whether or not to be greedy (chance is 1-e to use best action)
if (Util.RNG.NextDouble() < EGreedyThreshold)
{
// Explore a random action
return(RandomPlayoutBot.CreateRandomAction(state));
}
var action = new SabberStoneAction();
var stateClone = state.Game.Clone();
// Repeatedly exploit the highest (average) reward task that is available in this state
do
{
SabberStonePlayerTask selectedTask;
// Get the stats of the tasks currently available in this state
var availableTasks = stateClone.Game.CurrentPlayer.Options().Where(i => i.ZonePosition <= 0).Select(i => (SabberStonePlayerTask)i).ToList();
var availableTaskHashes = availableTasks.Select(i => i.GetHashCode()).ToList();
var availableStatistics = MASTTable.Where(i => availableTaskHashes.Contains(i.Key)).ToList();
// Find the task with the highest average value
var bestTask = availableStatistics.OrderByDescending(i => i.Value.AverageValue()).FirstOrDefault();
// If no best task was found, randomly choose an available task
if (bestTask.IsDefault())
{
var randomTask = availableTasks.RandomElementOrDefault();
// If we also can't randomly find a task, stop
if (randomTask == null)
{
break;
}
selectedTask = randomTask;
}
else
{
// Find all available tasks that have an average value similar to the best
var bestValue = bestTask.Value.AverageValue();
var compTasks = availableStatistics.Where(i =>
Math.Abs(i.Value.AverageValue() - bestValue) < AVThesis.Constants.DOUBLE_EQUALITY_TOLERANCE).ToList();
// Select one of the tasks
selectedTask = compTasks.RandomElementOrDefault().Value.Task;
}
// Add the task to the action we are building
action.AddTask(selectedTask);
// Process the task
stateClone.Process(selectedTask.Task);
// Continue until we have created a complete action, or the game has completed
} while (!action.IsComplete() && stateClone.Game.State != State.COMPLETE);
// Return the action we've created
return(action);
}
/// <summary>
/// Selects and action using the UCB1 algorithm.
/// </summary>
/// <param name="state">The game state.</param>
/// <returns><see cref="SabberStoneAction"/>.</returns>
private SabberStoneAction SelectUCB(SabberStoneState state)
{
var action = new SabberStoneAction();
var stateClone = state.Game.Clone();
// Repeatedly exploit the highest UCB-value task that is available in this state
do
{
SabberStonePlayerTask selectedTask;
// Get the stats of the tasks currently available in this state
var availableTasks = stateClone.Game.CurrentPlayer.Options().Where(i => i.ZonePosition <= 0).Select(i => (SabberStonePlayerTask)i).ToList();
var availableTaskHashes = availableTasks.Select(i => i.GetHashCode()).ToList();
var availableStatistics = MASTTable.Where(i => availableTaskHashes.Contains(i.Key)).ToList();
var totalVisits = availableStatistics.Sum(i => i.Value.Visits);
// Find the task with the highest UCB value
var bestTask = availableStatistics.OrderByDescending(i => i.Value.UCB(totalVisits, UCBConstantC)).FirstOrDefault();
// If no best task was found, randomly choose an available task
if (bestTask.IsDefault())
{
var randomTask = availableTasks.RandomElementOrDefault();
// If we also can't randomly find a task, stop
if (randomTask == null)
{
break;
}
selectedTask = randomTask;
}
else
{
// Find all available tasks that have an UCB value similar to the best
var bestValue = bestTask.Value.UCB(totalVisits, UCBConstantC);
var compTasks = availableStatistics.Where(i =>
Math.Abs(i.Value.UCB(totalVisits, UCBConstantC) - bestValue) < AVThesis.Constants.DOUBLE_EQUALITY_TOLERANCE).ToList();
// Select one of the tasks
selectedTask = compTasks.RandomElementOrDefault().Value.Task;
}
// Add the task to the action we are building
action.AddTask(selectedTask);
// Process the task
stateClone.Process(selectedTask.Task);
// Continue until we have created a complete action, or the game has completed
} while (!action.IsComplete() && stateClone.Game.State != State.COMPLETE);
// Return the action we've created
return(action);
}