void Start()
{
model = new MotionGenerator.Algorithm.Reinforcement.Models.Simple4Layer(
inputDimention: 1, outputDimention: 5, hiddenDimention: 30);
trainer = new MotionGenerator.Algorithm.Reinforcement.TemporalDifferenceQTrainer(
epsilon: 0.3f, qNetwork: model, historySize: 100000, discountRatio: 0.9f, actionDimention: 5,
replaySize: 32, rewardWeights: new[] { 1f });
state = Matrix <float> .Build.DenseDiagonal(1, 0);
lastReward = 0f;
}
private void CreateTrainer(int inputDimention)
{
if (_inputDimention != inputDimention)
{
_inputDimention = inputDimention;
_model.AlterInputDimention(inputDimention);
}
_trainer = new MotionGenerator.Algorithm.Reinforcement.TemporalDifferenceQTrainer(
epsilon: _randomActionProbability, qNetwork: _model,
historySize: _historySize, discountRatio: _discountRatio, actionDimention: Actions.Count,
replaySize: 32, alpha: _optimizerAlpha, rewardWeights: _soulWeights, optimizerType: _optimizerType,
enableRandomForgetting: _enableRandomForgetting,
initialHistory: _historySaveData != null
? _historySaveData.Select(x => x.Instantiate()).ToList()
: null
);
_historySaveData = null; //FIXME(kogaki): _historySaveDataをインスタンス変数に持たないようにしたい
}
void Start()
{
_logger = new TSVLogger();
if (IsMultiSoulModel)
{
model = new MotionGenerator.Algorithm.Reinforcement.Models.Simple4Layer(
inputDimention: 1, outputDimention: 5 * _soulWeights.Length, hiddenDimention: 30 * 2);
trainer = new MotionGenerator.Algorithm.Reinforcement.TemporalDifferenceQTrainer(
epsilon: 0.3f, qNetwork: model, historySize: 100000, discountRatio: 0.9f, actionDimention: 5,
replaySize: 32, rewardWeights: _soulWeights, alpha: 0.01f);
}
else
{
model = new MotionGenerator.Algorithm.Reinforcement.Models.Simple4Layer(
inputDimention: 1, outputDimention: 5, hiddenDimention: 30 * 2);
trainer = new MotionGenerator.Algorithm.Reinforcement.TemporalDifferenceQTrainer(
epsilon: 0.3f, qNetwork: model, historySize: 100000, discountRatio: 0.9f, actionDimention: 5,
replaySize: 32, rewardWeights: new[] { 1f }, alpha: 0.01f);
}
state = Matrix <float> .Build.DenseDiagonal(1, 0);
}