public double Sample(Agent agent, int horizon)
{
double reward = 0.0;
if (horizon == 0) {
return (int)reward;
}
else if (this.Type == ChanceNode) {
var percept = agent.GeneratePerceptAndUpdate();
int observation = percept.Item1;
int randomReward = percept.Item2;
if (!this.Children.ContainsKey(observation)) {//new node ->add it as decision node
this.Children[observation] = new MonteCarloSearchNode(DecisionNode);
}
MonteCarloSearchNode observationChild = this.Children[observation];
reward = randomReward + observationChild.Sample(agent, horizon-1);
}
else if (this.Visits == 0) //unvisited decision node or we have exceeded maximum tree depth
{
reward = agent.Playout(horizon);
// Console.WriteLine("from playout: reward ="+reward);
}
else { //Previously visited decision node
int actionNullable = this.SelectAction(agent);
int action = actionNullable;
agent.ModelUpdateAction(action);
if (!this.Children.ContainsKey(action)){ //this action is new chance child
this.Children[action]=new MonteCarloSearchNode(ChanceNode);
}
MonteCarloSearchNode actionChild = this.Children[action];
reward = actionChild.Sample(agent, horizon); //it is not clear if not horizon-1. (asks pyaixi)
}
double visitsDouble = this.Visits;
//Console.WriteLine("> {3} - {0}, {1}, {2}", this.mean, reward, (reward + (visitsDouble * this.mean) / (visitsDouble + 1.0)), visitsDouble);
this.Mean = (reward + (visitsDouble*this.Mean)) / (1.0 + visitsDouble);
this.Visits = this.Visits+1;
return reward;
}
//Interaction loop for interaction between agent and environment.
//This part is done in BrainSimulator in other version
// interaction begins with generating observation and reward from environment and giving it to agent
// agent then generates action and cycle repeats.
public static void InteractionLoop(Agent agent, AIXIEnvironment env, Dictionary<string, string> options)
{
Random rnd;
if (options.ContainsKey("random-seed"))
{
int seed;
int.TryParse(options["random-seed"], out seed);
rnd = new Random(seed);
}
else
{
rnd = new Random();
}
// Exploration = try random action
// probability will decay exponentially as exploreRate * exploreDecay ** round_number
var exploreRate = 0.0;
if (options.ContainsKey("exploration"))
{
exploreRate = Utils.MyToDouble(options["exploration"]);
}
var explore = exploreRate > 0;
var exploreDecay = 0.0;
if (options.ContainsKey("explore-decay"))
{
exploreDecay = Utils.MyToDouble(options["explore-decay"]);
}
Debug.Assert(0.0 <= exploreRate);
Debug.Assert(0.0 <= exploreDecay && exploreDecay <= 1.0);
//automatic halting after certain number of rounds
var terminateAge = 0;
if (options.ContainsKey("terminate-age"))
{
terminateAge = Convert.ToInt32(options["terminate-age"]);
}
var terminateCheck = terminateAge > 0;
Debug.Assert(0 <= terminateAge);
// when learning period passes, agent will stop changing/improving model and just use it.
var learningPeriod = 0;
if (options.ContainsKey("learning-period"))
{
learningPeriod = Convert.ToInt32(options["learning-period"]);
}
Debug.Assert(0 <= learningPeriod);
var cycle = 0;
while (!env.IsFinished)
{
if (terminateCheck && agent.Age > terminateAge)
{
break;
}
var cycleStartTime = DateTime.Now;
var observation = env.Observation;
var reward = env.Reward;
if (learningPeriod > 0 && cycle > learningPeriod)
{
explore = false;
}
//give observation and reward to agent.
agent.ModelUpdatePercept(observation, reward);
var explored = false;
int action;
if (explore && rnd.NextDouble() < exploreRate)
{
explored = true;
action = agent.GenerateRandomAction();
}
else
{
//get agents response to observation and reward
action = agent.Search();
}
//pass agent's action to environment
env.PerformAction(action);
agent.ModelUpdateAction(action);
var timeTaken = DateTime.Now - cycleStartTime;
Console.WriteLine("{0}:\t{1},{2},{3}\t{4},{5} \t{6},{7}\t>{8},{9}",
cycle, observation, reward, action,
explored, exploreRate,
agent.TotalReward, agent.AverageReward(),
timeTaken, agent.ModelSize()
);
if (explore)
{
exploreRate *= exploreDecay;
}
cycle += 1;
}
}
public int SelectAction(Agent agent)
{
Debug.Assert(agent.MaximumReward() != null, "this is weird place, - in selection action");
double exploreBias = (double)agent.Horizon * agent.MaximumReward().Value;
double explorationNumerator = this.ExplorationConstant * Math.Log(this.Visits);
int bestAction = -1;
double bestPriority = double.NegativeInfinity;
foreach (int action in agent.Environment.ValidActions) {
MonteCarloSearchNode node=null;
if (this.Children.ContainsKey(action)) {
node=this.Children[action];
}
double priority;
if (node == null || node.Visits == 0) {
// previously unexplored node
priority = this.UnexploredBias; //unexplored bias
}
else{
priority = node.Mean + exploreBias * Math.Sqrt(explorationNumerator / node.Visits);
}
if (priority > (bestPriority+Utils.RandomDouble(0, 0.001))){
bestAction=action;
bestPriority=priority;
}
}
return bestAction;
}