private float MSR(float[][] data, MFNN network)
{
int input_size = network.GetInputSize();
int output_size = network.GetOutputSize();
//Error Checking
Debug.Assert(data.Length > 0);
Debug.Assert(data[0].Length == input_size + output_size);
float msr = 0;
for (int i = 0; i < data.Length; i++)
{
float[] x_values = new float[input_size];
float[] t_values = new float[output_size];
Array.Copy(data[i], 0, x_values, 0, input_size);
Array.Copy(data[i], input_size, t_values, 0, output_size);
float[] y_values = network.Compute(x_values);
float sum = 0;
for (int j = 0; j < output_size; j++)
{
sum += (t_values[j] - y_values[j]) * (t_values[j] - y_values[j]);
}
msr += sum;
}
return(msr);
}
public void StartTraining()
{
last_q_values = new float[actions.Length];
abort_learning = false;
current_step = 0;
current_reward = 0;
particle_step = 0;
//Setup car components
car_camera = GetComponent <CarCamera>();
car_controller = GetComponent <CarController>();
car_body = transform.GetChild(0).GetComponent <Rigidbody>();
//Setup PSO
working_particle = 0;
particles = new MFNN[max_particles];
for (int i = 0; i < particles.Length; i++)
{
particles[i] = CreateMFNN();
}
particle_swarm = new PSO(particles);
network_target = CreateMFNN();
StartCoroutine(TakeStep());
}
public void StartTraining()
{
//Setup
action_index = 0;
current_step = 0;
learn_step = 0;
is_activated = true;
load_weights = false;
log = new List <string>();
last_q_values = new float[actions.Length];
//Setup memory
memory_index = 0;
network_memory = new MemoryStructure[memory_size];
//Setup networks
network_eval = new MFNN(
new int[] { 5, 20, actions.Length },
new ActivationType[] {
ActivationType.NONE,
ActivationType.ReLU,
ActivationType.ReLU
});
network_target = new MFNN(
new int[] { 5, 20, actions.Length },
new ActivationType[] {
ActivationType.NONE,
ActivationType.ReLU,
ActivationType.ReLU
});
//Setup epsilon
max_epsilon = 0.9f;
epsilon = 0.0f;
//Setup car components
car_camera = GetComponent <CarCamera>();
car_controller = GetComponent <CarController>();
car_body = gameObject.transform.GetChild(0).gameObject.GetComponent <Rigidbody>();
//Reset Car
car_body.transform.position = car_spawner.transform.position;
car_body.transform.rotation = car_spawner.transform.rotation;
car_body.velocity = Vector3.zero;
car_body.angularVelocity = Vector3.zero;
action_index = 0;
reset_step = current_step;
current_reward = 0;
//Start loop
StartCoroutine(DQNStep());
log.Add("Training Started");
}
private void Start()
{
MFNN network = new MFNN(new int[] { 4, 7, 3 }, new ActivationType[] {
ActivationType.NONE,
ActivationType.LOGISTIC_SIGMOID,
ActivationType.LOGISTIC_SIGMOID
});
int[] shuffle = ShuffleArray(IrisData.dataset.Length);
Debug.Log("Initial Error: " + MSR(IrisData.dataset, network));
int input_size = network.GetInputSize();
int output_size = network.GetOutputSize();
int r = 0;
while (r < 100)
{
for (int i = 0; i < IrisData.dataset.Length - 20; i++)
{
float[] x_values = new float[input_size];
float[] t_values = new float[output_size];
Array.Copy(IrisData.dataset[shuffle[i]], 0, x_values, 0, input_size);
Array.Copy(IrisData.dataset[shuffle[i]], input_size, t_values, 0, output_size);
float[] y_values = network.Compute(x_values);
float[] errors = new float[output_size];
for (int j = 0; j < output_size; j++)
{
errors[j] = t_values[j] - y_values[j];
}
network.UpdateWeights(errors, 0.01f, 0.0001f, 0.5f);
}
Debug.Log("Itr. " + r + " MSR: " + MSR(IrisData.dataset, network));
r++;
}
Debug.Log("Testing");
for (int i = IrisData.dataset.Length - 21; i < IrisData.dataset.Length; i++)
{
float[] x_values = new float[input_size];
float[] t_values = new float[output_size];
Array.Copy(IrisData.dataset[shuffle[i]], 0, x_values, 0, input_size);
Array.Copy(IrisData.dataset[shuffle[i]], input_size, t_values, 0, output_size);
float[] y_values = network.Compute(x_values);
int max1 = 0;
int max2 = 0;
for (int j = 0; j < t_values.Length; j++)
{
if (t_values[max1] < t_values[j])
{
max1 = j;
}
if (y_values[max2] < y_values[j])
{
max2 = j;
}
}
if (max1 == max2)
{
Debug.Log("GOOD");
}
else
{
Debug.Log("BAD");
}
}
}
