public void Initialize()
{
// Dimension Value Pairs:
sayWhat = World.NewDimensionValuePair("YourAction", "What do you want to do?");
// External Action Chunks:
sayCooperate = World.NewExternalActionChunk("Cooperate");
sayDefect = World.NewExternalActionChunk("Defect");
// placeholder
// GoalChunk salute = World.NewGoalChunk("Salute");
// GoalChunk bidFarewell = World.NewGoalChunk("Bid Farewell");
// WM Actions:
wmuacC = World.NewWorkingMemoryUpdateActionChunk("Remember my opponent cooperated");
wmuacD = World.NewWorkingMemoryUpdateActionChunk("Remember my opponent defected");
DeclarativeChunk dcoc = World.NewDeclarativeChunk("My opponent cooperated");
DeclarativeChunk dcod = World.NewDeclarativeChunk("My opponent defected");
wmuacC.Add(WorkingMemory.RecognizedActions.SET_RESET, dcoc);
wmuacD.Add(WorkingMemory.RecognizedActions.SET_RESET, dcod);
// Set up a two agent model (meaning two agents with the same setup, playing against each other)
Alice = World.NewAgent("Alice");
Bob = World.NewAgent("Bob");
// Simulating environment will determine inputs to each agent based on what each agent does..
// Feedback is determined by payoff matrix..
payoff = new int [2, 2, 2];
// Doing this the hard way. Could set this up all in-line above, but this makes the table
// more explicit in terms of how we want to use it.
// The payoff matrix here is called "Friend or Foe", about the simplest case
// indices mean: FOR-WHICH-AGENT, WHAT-ALICE-DOES, WHAT-BOB-DOES
payoff[_ALICE, _COOPERATE, _COOPERATE] = 1;
payoff[_ALICE, _COOPERATE, _DEFECT] = 0;
payoff[_ALICE, _DEFECT, _COOPERATE] = 2;
payoff[_ALICE, _DEFECT, _DEFECT] = 0;
payoff[_BOB, _COOPERATE, _COOPERATE] = 1;
payoff[_BOB, _COOPERATE, _DEFECT] = 2;
payoff[_BOB, _DEFECT, _COOPERATE] = 0;
payoff[_BOB, _DEFECT, _DEFECT] = 0;
maxpay = 2;
results = new int[_TRIALS, 2, 2];
// Set up a Q-learning Net =
// -- Eligibility Condition = True if "What do you want to do?" is in input, otherwise False
// -- Input = "My opponent cooperated", "My opponent defected", "What do you want to do?"
// -- Output = "I want to defect", "I want to cooperate"
//
// Also, RER is turned ON
QBPNetwork net_A = AgentInitializer.InitializeImplicitDecisionNetwork(Alice, QBPNetwork.Factory, QNetEC);
net_A.Input.Add(sayWhat);
net_A.Input.Add(sayCooperate);
net_A.Input.Add(sayDefect);
net_A.Output.Add(sayCooperate);
net_A.Output.Add(sayDefect);
Alice.Commit(net_A);
net_A.Parameters.LEARNING_RATE = 1;
Alice.ACS.Parameters.PERFORM_RER_REFINEMENT = true; // it's true by default anyway
Alice.ACS.Parameters.LEVEL_SELECTION_METHOD = ActionCenteredSubsystem.LevelSelectionMethods.COMBINED;
Alice.ACS.Parameters.LEVEL_SELECTION_OPTION = ActionCenteredSubsystem.LevelSelectionOptions.FIXED;
Alice.ACS.Parameters.FIXED_FR_LEVEL_SELECTION_MEASURE = 1;
Alice.ACS.Parameters.FIXED_BL_LEVEL_SELECTION_MEASURE = 1;
Alice.ACS.Parameters.FIXED_RER_LEVEL_SELECTION_MEASURE = 1;
Alice.ACS.Parameters.WM_UPDATE_ACTION_PROBABILITY = 1;
// Rules (2 rules) =
// Rule 1:
// -- Condition = "Your opponent cooperated"
// -- Action = Set "My opponent cooperated" in WM
// Rule 2:
// -- Condition = "Your opponent defected"
// -- Action = Set "My opponent defect" in WM
FixedRule ruleA1 = AgentInitializer.InitializeActionRule(Alice, FixedRule.Factory, wmuacC, FRSC);
FixedRule ruleA2 = AgentInitializer.InitializeActionRule(Alice, FixedRule.Factory, wmuacD, FRSC);
Alice.Commit(ruleA1);
Alice.Commit(ruleA2);
QBPNetwork net_B = AgentInitializer.InitializeImplicitDecisionNetwork(Bob, QBPNetwork.Factory, QNetEC);
net_B.Input.Add(sayWhat);
net_B.Input.Add(sayCooperate);
net_B.Input.Add(sayDefect);
net_B.Output.Add(sayCooperate);
net_B.Output.Add(sayDefect);
Bob.Commit(net_B);
// Use Weighted Combination
// NO partial match on TL
net_B.Parameters.LEARNING_RATE = 1;
Bob.ACS.Parameters.PERFORM_RER_REFINEMENT = true;
Bob.ACS.Parameters.LEVEL_SELECTION_METHOD = ActionCenteredSubsystem.LevelSelectionMethods.COMBINED;
Bob.ACS.Parameters.LEVEL_SELECTION_OPTION = ActionCenteredSubsystem.LevelSelectionOptions.FIXED;
Bob.ACS.Parameters.FIXED_FR_LEVEL_SELECTION_MEASURE = 1;
Bob.ACS.Parameters.FIXED_BL_LEVEL_SELECTION_MEASURE = 1;
Bob.ACS.Parameters.FIXED_RER_LEVEL_SELECTION_MEASURE = 1;
Bob.ACS.Parameters.WM_UPDATE_ACTION_PROBABILITY = 1;
FixedRule ruleB1 = AgentInitializer.InitializeActionRule(Bob, FixedRule.Factory, wmuacC, FRSC);
FixedRule ruleB2 = AgentInitializer.InitializeActionRule(Bob, FixedRule.Factory, wmuacD, FRSC);
Bob.Commit(ruleB1);
Bob.Commit(ruleB2);
// Initially using the same parameters for RER as Full Hello World
RefineableActionRule.GlobalParameters.SPECIALIZATION_THRESHOLD_1 = -.6;
RefineableActionRule.GlobalParameters.GENERALIZATION_THRESHOLD_1 = -.1;
RefineableActionRule.GlobalParameters.INFORMATION_GAIN_OPTION = RefineableActionRule.IGOptions.PERFECT;
/*
* Note -- What should be seems is that when you pass in "Your opponent…",
* the agent should return the "Do Nothing" external action
* (since it performed an internal WM action)..
* However, you can just ignore this either way..
*/
}
public void Initialize(Groups group)
{
World.Initialize();
John = World.NewAgent();
QBPNetwork idn = AgentInitializer.InitializeImplicitDecisionNetwork(John, QBPNetwork.Factory);
World.NewDimensionValuePair("Target P", target);
World.NewDimensionValuePair("Current P", target);
World.NewExternalActionChunk(target);
for (double i = 0; i < 12; i++)
{
if (World.GetDimensionValuePair("Target P", i) == null)
{
idn.Input.Add(World.NewDimensionValuePair("Target P", i));
idn.Input.Add(World.NewDimensionValuePair("Current P", i));
idn.Input.Add(World.NewExternalActionChunk(i));
idn.Output.Add(World.GetActionChunk(i));
}
else
{
idn.Input.Add(World.GetDimensionValuePair("Target P", i));
idn.Input.Add(World.GetDimensionValuePair("Current P", i));
idn.Input.Add(World.GetActionChunk(i));
idn.Output.Add(World.GetActionChunk(i));
}
}
foreach (double i in As)
{
World.NewDimensionValuePair("A", i);
}
foreach (double i in Bs)
{
World.NewDimensionValuePair("B", i);
}
foreach (double i in Cs)
{
World.NewDimensionValuePair("C", i);
}
switch (group)
{
case Groups.VERBALIZATION:
idn.Parameters.POSITIVE_MATCH_THRESHOLD = 1;
RefineableActionRule.GlobalParameters.POSITIVE_MATCH_THRESHOLD = 1;
RefineableActionRule.GlobalParameters.GENERALIZATION_THRESHOLD_1 = 1;
RefineableActionRule.GlobalParameters.SPECIALIZATION_THRESHOLD_1 = .5;
threshold_4 = .5;
break;
case Groups.MEMORY:
for (double i = 0; i < 12; i++)
{
ExternalActionChunk w = (ExternalActionChunk)World.GetActionChunk((double)rand.Next(12));
var p = World.GetDimensionValuePair("Current P", FactoryOutput(i, (double)w.LabelAsIComparable));
ExternalActionChunk w1 = (ExternalActionChunk)World.GetActionChunk(Math.Round((target + p.Value + NoiseOptions[rand.Next(3)]) / 2));
FixedRule mfr = AgentInitializer.InitializeActionRule(John, FixedRule.Factory, w1, MemoryGroup_SupportCalculator);
mfr.GeneralizedCondition.Add(p, true);
mfr.GeneralizedCondition.Add(w, true);
John.Commit(mfr);
}
goto default;
case Groups.SIMPLE_RULE:
for (double i = 0; i < 12; i++)
{
FixedRule sfr = AgentInitializer.InitializeActionRule(John, FixedRule.Factory, World.GetActionChunk(i), SimpleRule_SupportCalculator);
John.Commit(sfr);
}
goto default;
default:
idn.Parameters.LEARNING_RATE = .05;
idn.Parameters.DISCOUNT = .95;
John.ACS.Parameters.SELECTION_TEMPERATURE = .09;
idn.Parameters.POSITIVE_MATCH_THRESHOLD = 1;
RefineableActionRule.GlobalParameters.GENERALIZATION_THRESHOLD_1 = 2;
RefineableActionRule.GlobalParameters.SPECIALIZATION_THRESHOLD_1 = 1.2;
RefineableActionRule.GlobalParameters.POSITIVE_MATCH_THRESHOLD = 1;
threshold_4 = .2;
break;
}
RefineableActionRule.GlobalParameters.INFORMATION_GAIN_OPTION = RefineableActionRule.IGOptions.PERFECT;
John.Commit(idn);
}