/// <summary>
/// 生成可以测试的动作计划集：从动作记忆中找到的行动计划，加上新补充的一些
/// </summary>
/// <param name="plans">从动作记忆中找到的行动计划</param>
/// <param name="time"></param>
/// <returns></returns>
private List <ActionPlan> checkActionPlansFull(List <ObservationHistory.ActionRecord> actionRecords, int time)
{
List <List <double> > actionSets = CreateTestActionSet(Session.instinctActionHandler(net, time));
ActionPlan[] r = new ActionPlan[actionSets.Count];
for (int i = 0; i < actionSets.Count; i++)
{
ActionPlan plan = null;
String judgeType = ActionPlan.JUDGE_INFERENCE;
if (i == 0)
{
judgeType = ActionPlan.JUDGE_INSTINCT;
}
else if (actionSets[i][0] == 0.5)
{
judgeType = ActionPlan.JUDGE_MAINTAIN;
}
ObservationHistory.ActionRecord record = actionRecords.FirstOrDefault(p => p.actions[0] == actionSets[i][0]);
if (record == null)
{
plan = ActionPlan.CreateActionPlan(net, actionSets[i], time, judgeType, "");
}
else
{
plan = ActionPlan.CreateActionPlan(net, record.actions, time, judgeType, "");
plan.evaulation = record.evaluation;
}
r[i] = plan;
}
return(r.ToList());
}
/// <summary>
/// 对当前行为进行推理，预测其未来评估
/// </summary>
/// <returns></returns>
private double forcastActionPlan()
{
if (this.net.actionPlanChain.Length <= 0)
{
return(double.NaN);
}
//寻找当前场景，如果继续维持当前行为，是否会导致负评估
//由于漫游空间大，碰找到的几率不大
List <ObservationHistory.ActionRecord> actionRecords = this.net.actionMemory.FindMatchActionPlans();
if (actionRecords != null && actionRecords.Count > 0)
{
ObservationHistory.ActionRecord record = actionRecords.FirstOrDefault(r => r.actions[0] == 0.5);
if (record != null)
{
return(record.evaluation);
}
}
//执行推理操作
List <Vector> observations = net.GetReceoptorValues();
observations = net.ReplaceWithAction(observations, ActionPlan.MaintainAction);
return(forcastActionPlan(observations));
}
private double forcastActionPlan(List <Vector> observation)
{
int maxinferencecount = Session.GetConfiguration().evaluation.policy.init_plan_depth;
for (int n = 1; n <= maxinferencecount; n++)
{
List <Vector> nextObservation = net.forward_inference(observation);
if (nextObservation == null)
{
break;
}
List <ObservationHistory.ActionRecord> actionRecords = this.net.actionMemory.FindMatchActionPlans(nextObservation);
if (actionRecords != null && actionRecords.Count > 0)
{
ObservationHistory.ActionRecord record = actionRecords.FirstOrDefault(r => r.actions[0] == 0.5);
if (record != null)
{
return(record.evaluation > 0 ? n + record.evaluation : -1 * n + record.evaluation);
}
}
observation = nextObservation;
net.ReplaceMaintainAction(observation);
}
return(double.NaN);
}
