public void 遺伝によるコピーで作られたDecisionMakerは完全に同じ動きをする()
{
// Init parent
var parentDecisionMaker = new ReinforcementDecisionMaker();
var actions = new List <IAction>();
for (int i = 0; i < 100; i++)
{
actions.Add(LocomotionAction.GoStraight(i.ToString()));
}
parentDecisionMaker.Init(actions);
parentDecisionMaker.DecideAction(_dummyStates[0], forceRandom: false, forceMax: true);
// Init child
var childDecisionMaker = new ReinforcementDecisionMaker();
childDecisionMaker.Init(parentDecisionMaker);
// Assertion
foreach (var state in _dummyStates)
{
Assert.AreEqual(
parentDecisionMaker.DecideAction(state, forceRandom: false, forceMax: true)
.Name, // choose from 100 choice
childDecisionMaker.DecideAction(state, forceRandom: false, forceMax: true).Name
);
}
}
public void SubDecisionMakerを使って意思決定ができ遺伝もできる()
{
// Init parent
var parentDecisionMaker = new ReinforcementDecisionMaker();
var actions = new List <IAction>();
actions.AddRange(LocomotionAction.EightDirections());
for (int i = 0; i < 100; i++)
{
actions.Add(
new SubDecisionMakerAction(new FollowPointDecisionMaker(State.BasicKeys.RelativeFoodPosition)));
}
parentDecisionMaker.Init(actions);
parentDecisionMaker.DecideAction(_dummyStates[0], forceRandom: false, forceMax: true);
// Init child
var childDecisionMaker = new ReinforcementDecisionMaker();
childDecisionMaker.Init(parentDecisionMaker);
// Assertion
foreach (var state in _dummyStates)
{
var parentAction = parentDecisionMaker.DecideAction(state, forceRandom: false, forceMax: true);
var childAction = childDecisionMaker.DecideAction(state, forceRandom: false, forceMax: true);
Assert.AreEqual(
parentAction.Name,
childAction.Name
);
// SubDecisionMakerActionは外に出ない
Assert.IsInstanceOf <LocomotionAction>(parentAction);
Assert.IsInstanceOf <LocomotionAction>(childAction);
}
}
public void デシリアライズ後RestoreされたものはSubDMも含めて同じDecisionをする()
{
// Init parent
var parentDecisionMaker = new ReinforcementDecisionMaker();
var actions = new List <IAction>();
actions.Add(new SubDecisionMakerAction(new FollowPointDecisionMaker(State.BasicKeys.RelativeFoodPosition)));
actions.AddRange(LocomotionAction.EightDirections());
parentDecisionMaker.Init(actions);
parentDecisionMaker.DecideAction(_dummyStates[0], forceRandom: false, forceMax: true);
var saveDataClone = EditorTestExtensions.DeepCloneByMsgPack(parentDecisionMaker.Save());
var decisionMakerClone = saveDataClone.Instantiate() as ReinforcementDecisionMaker;
decisionMakerClone.Restore(actions);
// Random
foreach (var state in _dummyStates)
{
Assert.AreEqual(
parentDecisionMaker.DecideAction(state, forceRandom: false, forceMax: true)
.Name,
decisionMakerClone.DecideAction(state, forceRandom: false, forceMax: true).Name
);
}
// Force SubDM
foreach (var state in _dummyStates)
{
Assert.AreEqual(
parentDecisionMaker.DecideAction(state, forceRandom: false, forceMax: true, forceAction: 0)
.Name, // choose from 100 choice
decisionMakerClone.DecideAction(state, forceRandom: false, forceMax: true, forceAction: 0).Name
);
}
}
public void デシリアライズ後Restoreされたものは経験を引き継ぐ()
{
// Init parent
var parentDecisionMaker = new ReinforcementDecisionMaker();
var actions = new List <IAction>();
for (int i = 0; i < 100; i++)
{
actions.Add(LocomotionAction.GoStraight(i.ToString()));
}
parentDecisionMaker.Init(actions);
parentDecisionMaker.DecideAction(_dummyStates[0], forceRandom: false, forceMax: true);
parentDecisionMaker.Feedback(new List <float>()
{
3f
});
parentDecisionMaker.DecideAction(_dummyStates[0], forceRandom: false, forceMax: true);
var saveDataClone = EditorTestExtensions.DeepCloneByMsgPack(parentDecisionMaker.Save());
var decisionMakerClone = saveDataClone.Instantiate() as ReinforcementDecisionMaker;
var originalTrainer = TestHelper.GetFieldValue(parentDecisionMaker, "_trainer") as TemporalDifferenceQTrainer;
var cloneTrainer = TestHelper.GetFieldValue(decisionMakerClone, "_trainer") as TemporalDifferenceQTrainer;
// Assertion
Assert.AreEqual(originalTrainer.GetHistorySaveData().Count, 1);
Assert.AreEqual(cloneTrainer.GetHistorySaveData().Count, 1);
Assert.AreEqual(originalTrainer.GetHistorySaveData()[0].Instantiate().State, cloneTrainer.GetHistorySaveData()[0].Instantiate().State);
}
private float HeuristicPenaltyWalking(State lastState, State nowState)
{
var penalty = 1f;
var movement = LocomotionAction.GetLastMovement(lastState, nowState);
var targetStepSize = nowState.GetAsFloat(State.BasicKeys.BodyScale);
var overDistance = movement.magnitude - targetStepSize;
if (0f < overDistance)
{
// オーバーランはペナルティ
var overPenalty = targetStepSize - overDistance;
if (overPenalty > 0f)
{
penalty *= overPenalty;
}
else
{
// マイナスだと逆方向になるので0
penalty = 0f;
}
}
return(penalty);
}
public void デシリアライズ後Restoreされたものは同じDecisionをする()
{
// Init parent
var parentDecisionMaker = new ReinforcementDecisionMaker();
var actions = new List <IAction>();
for (int i = 0; i < 100; i++)
{
actions.Add(LocomotionAction.GoStraight(i.ToString()));
}
parentDecisionMaker.Init(actions);
parentDecisionMaker.DecideAction(_dummyStates[0], forceRandom: false, forceMax: true);
var saveDataClone = EditorTestExtensions.DeepCloneByMsgPack(parentDecisionMaker.Save());
var decisionMakerClone = saveDataClone.Instantiate() as ReinforcementDecisionMaker;
// Assertion
foreach (var state in _dummyStates)
{
Assert.AreEqual(
parentDecisionMaker.DecideAction(state, forceRandom: false, forceMax: true)
.Name, // choose from 100 choice
decisionMakerClone.DecideAction(state, forceRandom: false, forceMax: true).Name
);
}
}
public DummyRemoteDecisionMaker()
{
Fetcher = new DummyFetcher();
Actions = new List <IAction>()
{
LocomotionAction.GoStraight("test")
};
}
public void 左に餌があるときはちゃんとそちらに向かう()
{
var decisionMaker = createDummy();
var tmpState = new State();
for (var y = -1; y <= 1; y++)
{
tmpState [State.BasicKeys.RelativeFoodPosition] = new DenseVector(new double[] { -2, y * 100, 0.5f });
var action = decisionMaker.DecideAction(tmpState);
Assert.AreEqual(
LocomotionAction.GoLeft().Name,
action.Name
);
}
}
public void 反対側にちゃんと向かう()
{
var decisionMaker = createDummy(isNegative: true);
var tmpState = new State();
for (var y = -1; y <= 1; y++)
{
tmpState[State.BasicKeys.RelativeFoodPosition] = new DenseVector(new double[] { 0.5f, y * 100, 2 });
var action = decisionMaker.DecideAction(tmpState);
Assert.AreEqual(
LocomotionAction.GoBack().Name,
action.Name
);
}
}
public void 左にエネルギーがあるときはちゃんとそちらに向かう()
{
var decisionMaker = createDummy();
var tmpState = TempState();
for (var y = -1; y <= 1; y++)
{
tmpState[State.BasicKeys.TotalFoodEnergyEachDirection][6] = 1f;
var action = decisionMaker.DecideAction(tmpState);
Assert.AreEqual(
LocomotionAction.GoLeft().Name,
action.Name
);
}
}
public void 反対側をちゃんと選ぶ()
{
var decisionMaker = createDummy(isNegative: true);
var tmpState = TempState();
for (var y = -1; y <= 1; y++)
{
tmpState[State.BasicKeys.TotalFoodEnergyEachDirection][0] = 1f;
var action = decisionMaker.DecideAction(tmpState);
Assert.AreEqual(
LocomotionAction.GoBack().Name,
action.Name
);
}
}
public void RewardTest()
{
//Arrange
var action = LocomotionAction.GoStraight("test");
var lastPosition = new Vector3(1, 2, 3);
var nowPosition = new Vector3(3, 4, 5);
var lastState = new State();
lastState[State.BasicKeys.Position] =
new DenseVector(new double[] { lastPosition.x, lastPosition.y, lastPosition.z });
lastState[State.BasicKeys.Rotation] = new DenseVector(new double[] { 0, 0, 0, 0 });
var nowState = new State();
nowState[State.BasicKeys.Position] =
new DenseVector(new double[] { nowPosition.x, nowPosition.y, nowPosition.z });
//Act
var reward = action.Reward(lastState, nowState);
//Assert
Assert.AreEqual(reward, 2.0f);
}
public LocomotionAction(LocomotionAction src) : base(src)
{
Direction = src.Direction;
}
public static List <IAction> EightDirections()
{
return(new List <IAction>(LocomotionAction.EightDirectionsAsLocomotionActions()
.Select(locomotionAction => new WalkingLocomotionAction(locomotionAction)).ToList()));
}
public WalkingLocomotionAction(LocomotionAction baseAction) : base("Walking" + baseAction.Name)
{
BaseAction = baseAction;
}
