public void AfterTrainingJob(NeuralBurst trainee, float dt)
{
var traineeBody = trainee.cachedRigidbody;
var brake = traineeBody.velocity * -trainee.state2[3];
var force = new Vector3(
trainee.state2[0],
trainee.state2[1],
trainee.state2[2]
) * 1000.0f;
//if (float.IsNaN(force.x)) force.x = 0.0f;
//if (float.IsNaN(force.y)) force.y = 0.0f;
//if (float.IsNaN(force.z)) force.z = 0.0f;
var apply = (brake + force) * dt;
traineeBody.AddForce(apply, ForceMode.Acceleration);
if (showDebugLines)
{
//Debug.DrawLine(new Vector3(trainee.state0[0], trainee.state0[1], trainee.state0[2]), new Vector3(trainee.state0[3], trainee.state0[4], trainee.state0[5]), new Color(trainee.state2[0] * 2.0f, trainee.state2[1] * 2.0f, trainee.state2[2] * 2.0f));
//Debug.DrawLine(new Vector3(trainee.state0[0], trainee.state0[1], trainee.state0[2]), new Vector3(trainee.state0[6], trainee.state0[7], trainee.state0[8]), Color.white);
var accuracy = Vector3.Dot(force.normalized, (target.position - traineeBody.position).normalized);
Debug.DrawLine(trainee.transform.position, trainee.transform.position + force * 0.01f, Color.Lerp(Color.white, Color.red, accuracy));
}
}
public void CalculateResult(NeuralBurst trainee, TrainingResult result)
{
var ofs = trainee.transform.position - target.transform.position;
var lsq = ofs.sqrMagnitude;
var len = 0.0f;
if (lsq > 0.001f)
{
len = ofs.magnitude;
}
result.score = len;
}
public NeuralBurst.StepJob PrepareAndQueueJobStep(NeuralBurst trainee, float dt)
{
var traineeBody = trainee.cachedRigidbody;
trainee.state0[0] = target.transform.position.x - trainee.transform.position.x;
trainee.state0[1] = target.transform.position.y - trainee.transform.position.y;
trainee.state0[2] = target.transform.position.z - trainee.transform.position.z;
trainee.state0[3] = traineeBody.velocity.x;
trainee.state0[4] = traineeBody.velocity.y;
trainee.state0[5] = traineeBody.velocity.z;
return(trainee.QueueJobStep());
}
public EvolveJob QueueJobEvolve(uint seed, NeuralBurst from, float evolveRate)
{
var job = new EvolveJob()
{
Seed = seed,
EvolveRate = evolveRate,
Layer0 = this.Layer0,
Layer1 = this.Layer1,
Layer2 = this.Layer2,
from_bias0 = from.bias0,
from_bias1 = from.bias1,
from_bias2 = from.bias2,
from_weights10 = from.weights10,
from_weights21 = from.weights21,
bias0 = this.bias0,
bias1 = this.bias1,
bias2 = this.bias2,
weights10 = this.weights10,
weights21 = this.weights21,
};
return(job);
}
public void Evolve(NeuralBurst from, float x)
{
for (int i = 0; i < Layer0; ++i)
{
bias0[i] = UnityEngine.Random.Range(0.0f, 1.0f) < x ? from.bias0[i] : bias0[i];
}
for (int i = 0; i < Layer1; ++i)
{
bias1[i] = UnityEngine.Random.Range(0.0f, 1.0f) < x ? from.bias1[i] : bias1[i];
}
for (int i = 0; i < Layer2; ++i)
{
bias2[i] = UnityEngine.Random.Range(0.0f, 1.0f) < x ? from.bias2[i] : bias2[i];
}
for (int i = 0; i < Layer0 * Layer1; ++i)
{
weights10[i] = UnityEngine.Random.Range(0.0f, 1.0f) < x ? from.weights10[i] : weights10[i];
}
for (int i = 0; i < Layer1 * Layer2; ++i)
{
weights21[i] = UnityEngine.Random.Range(0.0f, 1.0f) < x ? from.weights21[i] : weights21[i];
}
}
public void LerpTowards(NeuralBurst from, float x)
{
for (int i = 0; i < Layer0; ++i)
{
bias0[i] = Mathf.Lerp(bias0[i], from.bias0[i], x);
}
for (int i = 0; i < Layer1; ++i)
{
bias1[i] = Mathf.Lerp(bias1[i], from.bias1[i], x);
}
for (int i = 0; i < Layer2; ++i)
{
bias2[i] = Mathf.Lerp(bias2[i], from.bias2[i], x);
}
for (int i = 0; i < Layer0 * Layer1; ++i)
{
weights10[i] = Mathf.Lerp(weights10[i], from.weights10[i], x);
}
for (int i = 0; i < Layer1 * Layer2; ++i)
{
weights21[i] = Mathf.Lerp(weights21[i], from.weights21[i], x);
}
}
public IEnumerator RunTrainingLoop()
{
Debug.LogFormat("NeuralTrainerEvolution_Follow.RunTrainingLoop(): running...");
var cycle = 0;
bestScore = float.MaxValue;
bestTrainee = (NeuralBurst)null;
var stepJobs = new List <JobHandle>(trainees.Count);
var mutateJobs = new List <JobHandle>(trainees.Count);
var evolveJobs = new List <JobHandle>(trainees.Count);
while (true)
{
ClearResults();
var worldOffset = Vector3.zero;
if (trainingPhase >= 5)
{
worldOffset = Random.onUnitSphere * Random.Range(10.0f, 20.0f);
}
for (int i = 0; i < trainees.Count; ++i)
{
var trainee = trainees[i];
var result = results[i];
var direction = Random.onUnitSphere;
direction.y = 0.0f;
direction = direction.normalized;
var position = new Vector3(
Random.Range(-10.0f, 10.0f),
Random.Range(-10.0f, 10.0f),
Random.Range(-10.0f, 10.0f));
trainee.transform.position = position + worldOffset;
trainee.transform.rotation = Quaternion.LookRotation(Random.onUnitSphere);
trainee.GetComponent <Renderer>().material.color = trainee.GenerateColor();
}
switch (trainingPhase)
{
case 0:
target.transform.position = Vector3.zero;
break;
case 1:
{
var position = Random.onUnitSphere * 5.0f;
position.y = 0.0f;
target.transform.position = position;
}
break;
case 2:
target.transform.position = Random.onUnitSphere * 5.0f;
break;
case 3:
target.transform.position = Random.onUnitSphere * Random.Range(0.0f, 10.0f);
break;
case 4:
target.transform.position = Random.onUnitSphere * Random.Range(0.0f, 10.0f);
break;
case 5:
target.transform.position = Random.onUnitSphere * Random.Range(0.0f, 10.0f) + worldOffset;
break;
}
var counter = 0.0f;
while (counter < cycleTime)
{
if (hyper)
{
Physics.autoSimulation = false;
Time.timeScale = 1.0f * (float)hyperSpeed;
TickPhase(Time.fixedDeltaTime);
for (int h = 0; h < hyperSpeed; ++h)
{
for (int i = 0; i < trainees.Count; ++i)
{
var trainee = trainees[i];
var result = results[i];
var dt = Time.fixedDeltaTime;
var job = PrepareAndQueueJobStep(trainee, dt);
stepJobs.Add(job.Schedule());
if ((i % 8) == 0)
{
JobHandle.ScheduleBatchedJobs();
}
}
for (int i = 0; i < trainees.Count; ++i)
{
var trainee = trainees[i];
var handle = stepJobs[i];
handle.Complete();
AfterTrainingJob(trainee, Time.fixedDeltaTime);
}
stepJobs.Clear();
Physics.Simulate(Time.fixedDeltaTime);
}
counter += Time.fixedDeltaTime * Time.timeScale;
yield return(null);
}
else
{
Physics.autoSimulation = true;
Time.timeScale = 1.0f;
TickPhase(Time.fixedDeltaTime);
UnityEngine.Profiling.Profiler.BeginSample("QueueTrainingJobs");
for (int i = 0; i < trainees.Count; ++i)
{
var trainee = trainees[i];
var result = results[i];
var job = PrepareAndQueueJobStep(trainee, Time.fixedDeltaTime);
stepJobs.Add(job.Schedule());
if ((i % 8) == 0)
{
JobHandle.ScheduleBatchedJobs();
}
}
UnityEngine.Profiling.Profiler.EndSample();
UnityEngine.Profiling.Profiler.BeginSample("AfterTrainingJobs");
for (int i = 0; i < trainees.Count; ++i)
{
var trainee = trainees[i];
var handle = stepJobs[i];
handle.Complete();
AfterTrainingJob(trainee, Time.fixedDeltaTime);
}
stepJobs.Clear();
UnityEngine.Profiling.Profiler.EndSample();
counter += Time.fixedDeltaTime;
yield return(new WaitForFixedUpdate());
}
}
for (int i = 0; i < trainees.Count; ++i)
{
var trainee = trainees[i];
var result = results[i];
CalculateResult(trainee, result);
}
results.Sort(CompareResults);
var winner = results[0];
if (winner.score < bestScore)
{
Debug.LogFormat("NeuralTrainerEvolution_Follow.RunTrainingLoop(): cycle {0} +++ new winner: {1}: score: {2}", cycle, winner.trainee.name, winner.score);
bestScore = winner.score;
bestTrainee = winner.trainee;
bestScore += 0.1f;
}
else
{
Debug.LogFormat("NeuralTrainerEvolution_Follow.RunTrainingLoop(): cycle {0} === old winner: {1}: score: {2}", cycle, bestTrainee.name, bestScore);
bestScore += 1.0f;
}
for (int i = 0; i < trainees.Count; ++i)
{
var seed = (uint)(Time.frameCount * (i + 1));
if (trainees[i] == bestTrainee)
{
continue;
}
var jobMutate = trainees[i].QueueJobMutate(seed, mutateRate);
var jobEvolve = trainees[i].QueueJobEvolve(seed, bestTrainee, evolveRate);
var jobHandleMutate = jobMutate.Schedule();
var jobHandleEvolve = jobEvolve.Schedule(jobHandleMutate);
mutateJobs.Add(jobHandleMutate);
evolveJobs.Add(jobHandleEvolve);
if ((i % 8) == 0)
{
JobHandle.ScheduleBatchedJobs();
}
}
// this shouldnt be needed but getting dependency error
for (int i = 0; i < mutateJobs.Count; ++i)
{
mutateJobs[i].Complete();
}
mutateJobs.Clear();
for (int i = 0; i < evolveJobs.Count; ++i)
{
evolveJobs[i].Complete();
}
evolveJobs.Clear();
for (int i = 0; i < trainees.Count; ++i)
{
trainees[i].cachedRigidbody.isKinematic = true;
}
yield return(null);
for (int i = 0; i < trainees.Count; ++i)
{
trainees[i].cachedRigidbody.isKinematic = false;
}
cycle++;
yield return(null);
}
}