Determinet

?? Be sure to check out the NuGet pacakge: https://www.nuget.org/packages/NTDLS.Determinet

Determinet is versatile multilayer perception neural network designed for extendibility and genetic-style mutations to allow forward propagation of the network variants.

Below is a simple example of using the network to navigate a maze or other obstacles for a simple simulation. You can find more advanced examples as well as working models of this in the AIVolution project: https://github.com/NTDLS/AIVolution

If you are in a fighting mood, you can also battle it out against some of these trained models in the Space Flight Shooter Game: https://github.com/NTDLS/NebulaSiege

public enum AIInputs
{
    DistanceFromObstacle,
    AngleToObstacleInDecimalDegrees
}

public enum AIOutputs
{
    MoveAway,
    AdjustSpeed
}

static void Main()
{
    TrainAndSaveModel("./TestHarness.json");

    //Note that if you want to use the model in different threads, you will need to
    //  make a clone since the values of the input nodes are altered when making decisions.
    //  Fortunately, this can be easily accomplished with a call to Clone();
    var network = LoadSavedModel("./TestHarness.json");

    var decidingFactors = GatherInputs(); //Get decision inputs.

    var decisions = network.FeedForward(decidingFactors); //Make decisions.

    //Handle the speed decisions.
    var shouldAdjustSpeed = decisions.Get(AIOutputs.AdjustSpeed);
    if (shouldAdjustSpeed > 0.8)
    {
        //Adjust speed up.
    }
    else if (shouldAdjustSpeed < 0.2)
    {
        //Adjust speed down.
    }

    //Handle the heading direction.
    var shouldMoveAway = decisions.Get(AIOutputs.MoveAway);
    if (shouldMoveAway > 0.9)
    {
        //Change heading. Maybe just turn around completely?
    }
}

/// <summary>
/// Get the input values we need to make a decision.
/// </summary>
/// <returns></returns>
static DniNamedInterfaceParameters GatherInputs()
{
    //Here we are just using some dummy values, in this hypothetical situation
    //  we would be getting the distance from a wall and the angle to it.

    double idealMaxDistance = 1000;
    double distanceFromObstacle = 500;

    double percentageOfCloseness = (distanceFromObstacle / idealMaxDistance);
    double angleToObstacleInDecimalDegrees = 0.8;

    var aiParams = new DniNamedInterfaceParameters();
    aiParams.Set(AIInputs.DistanceFromObstacle, percentageOfCloseness);
    aiParams.Set(AIInputs.AngleToObstacleInDecimalDegrees, angleToObstacleInDecimalDegrees);
    return aiParams;
}

static DniNeuralNetwork LoadSavedModel(string fileName)
{
    var network = DniNeuralNetwork.LoadFromFile(fileName);

    if (network == null)
    {
        throw new Exception("Failed to load the network from file.");
    }

    return network;
}

static void TrainAndSaveModel(string fileName)
{
    var Network = new DniNeuralNetwork
    {
        LearningRate = 0.01
    };

    //Add input layer
    Network.Layers.AddInput(ActivationType.LeakyReLU,
        new object[] {
                AIInputs.DistanceFromObstacle,
                AIInputs.AngleToObstacleInDecimalDegrees
        });

    //Add a intermediate "hidden" layer. You can add more if you like.
    Network.Layers.AddIntermediate(ActivationType.Sigmoid, 8);

    //Add the output layer.
    Network.Layers.AddOutput(
        new object[] {
                AIOutputs.MoveAway,
                AIOutputs.AdjustSpeed
        });

    //Train the model with some input scenarios. Look at TrainingScenerio() and TrainingDecision()
    //  to see that these ominous looking numbers are actualy just named inouts. Its pretty simple really.
    for (int epoch = 0; epoch < 5000; epoch++)
    {
        //Very close to observed object, slow way down and get away
        Network.BackPropagate(TrainingScenerio(0, 0), TrainingDecision(1, 0));
        Network.BackPropagate(TrainingScenerio(0, -1), TrainingDecision(1, 0));
        Network.BackPropagate(TrainingScenerio(0, 1), TrainingDecision(1, 0));
        Network.BackPropagate(TrainingScenerio(0, 0.5), TrainingDecision(1, 0));
        Network.BackPropagate(TrainingScenerio(0, -0.5), TrainingDecision(1, 0));

        //Pretty close to observed object, slow down a bit and get away.
        Network.BackPropagate(TrainingScenerio(0.25, 0), TrainingDecision(1, 0.2));
        Network.BackPropagate(TrainingScenerio(0.25, -1), TrainingDecision(1, 0.2));
        Network.BackPropagate(TrainingScenerio(0.25, 1), TrainingDecision(1, 0.2));
        Network.BackPropagate(TrainingScenerio(0.25, 0.5), TrainingDecision(1, 0.2));
        Network.BackPropagate(TrainingScenerio(0.25, -0.5), TrainingDecision(1, 0.2));

        //Very far from observed object, speed up and maintain heading.
        Network.BackPropagate(TrainingScenerio(1, 0), TrainingDecision(0, 1));
        Network.BackPropagate(TrainingScenerio(1, -1), TrainingDecision(0, 1));
        Network.BackPropagate(TrainingScenerio(1, 1), TrainingDecision(0, 1));
        Network.BackPropagate(TrainingScenerio(1, 0.5), TrainingDecision(0, 1));
        Network.BackPropagate(TrainingScenerio(1, -0.5), TrainingDecision(0, 1));

        //Pretty far from observed object, maintain heading but don't change speed.
        Network.BackPropagate(TrainingScenerio(0.75, 0), TrainingDecision(0, 0.5));
        Network.BackPropagate(TrainingScenerio(0.75, -1), TrainingDecision(0, 0.5));
        Network.BackPropagate(TrainingScenerio(0.75, 1), TrainingDecision(0, 0.5));
        Network.BackPropagate(TrainingScenerio(0.75, 0.5), TrainingDecision(0, 0.5));
        Network.BackPropagate(TrainingScenerio(0.75, -0.5), TrainingDecision(0, 0.5));
    }

    static DniNamedInterfaceParameters TrainingScenerio(double distanceFromObstacle, double angleToObstacleInDecimalDegrees)
    {
        var param = new DniNamedInterfaceParameters();
        param.Set(AIInputs.DistanceFromObstacle, distanceFromObstacle);
        param.Set(AIInputs.AngleToObstacleInDecimalDegrees, angleToObstacleInDecimalDegrees);
        return param;
    }

    static DniNamedInterfaceParameters TrainingDecision(double moveAway, double adjustSpeed)
    {
        var param = new DniNamedInterfaceParameters();

        param.Set(AIOutputs.MoveAway, moveAway);
        param.Set(AIOutputs.AdjustSpeed, adjustSpeed);
        return param;
    }

    //Save the network to a file. This is only done here for examples sake.
    Network.Save(fileName);
}

License

Apache-2.0