private void AgentCrossover(int playerIndex)
{
List <BrainGenome> newGenBrainGenomeList = new List <BrainGenome>(); // new population!
FitnessManager fitnessManager = teamsConfig.playersList[playerIndex].fitnessManager;
TrainingSettingsManager trainingSettingsManager = teamsConfig.playersList[playerIndex].trainingSettingsManager;
// Keep top-half peformers + mutations:
for (int x = 0; x < teamsConfig.playersList[playerIndex].agentGenomeList.Count; x++)
{
if (x == 0)
{
BrainGenome parentGenome = teamsConfig.playersList[playerIndex].agentGenomeList[fitnessManager.rankedIndicesList[x]].brainGenome;
//parentGenome.index = 0;
newGenBrainGenomeList.Add(parentGenome);
}
else
{
BrainGenome newBrainGenome = new BrainGenome();
int parentIndex = fitnessManager.GetAgentIndexByLottery();
BrainGenome parentGenome = teamsConfig.playersList[playerIndex].agentGenomeList[parentIndex].brainGenome;
newBrainGenome.SetToMutatedCopyOfParentGenome(parentGenome, trainingSettingsManager);
newGenBrainGenomeList.Add(newBrainGenome);
}
}
for (int i = 0; i < teamsConfig.playersList[playerIndex].agentGenomeList.Count; i++)
{
teamsConfig.playersList[playerIndex].agentGenomeList[i].brainGenome = newGenBrainGenomeList[i];
}
}
private void EnvironmentCrossover()
{
List <EnvironmentGenome> newGenGenomeList = new List <EnvironmentGenome>(); // new population!
FitnessManager fitnessManager = teamsConfig.environmentPopulation.fitnessManager;
TrainingSettingsManager trainingSettingsManager = teamsConfig.environmentPopulation.trainingSettingsManager;
float mutationChance = trainingSettingsManager.mutationChance;
float mutationStepSize = trainingSettingsManager.mutationStepSize;
// Keep top-half peformers + mutations:
for (int x = 0; x < teamsConfig.environmentPopulation.environmentGenomeList.Count; x++)
{
if (false) //x == 0) {
// Top performer stays
{
EnvironmentGenome parentGenome = teamsConfig.environmentPopulation.environmentGenomeList[fitnessManager.rankedIndicesList[x]];
parentGenome.index = 0;
newGenGenomeList.Add(parentGenome);
}
else
{
int parentIndex = fitnessManager.GetAgentIndexByLottery();
EnvironmentGenome parentGenome = teamsConfig.environmentPopulation.environmentGenomeList[parentIndex];
EnvironmentGenome newGenome = parentGenome.BirthNewGenome(parentGenome, newGenGenomeList.Count, teamsConfig.challengeType, mutationChance, mutationStepSize);
newGenGenomeList.Add(newGenome);
}
}
for (int i = 0; i < teamsConfig.environmentPopulation.environmentGenomeList.Count; i++)
{
teamsConfig.environmentPopulation.environmentGenomeList[i] = newGenGenomeList[i];
}
}
public void StartNewTrainingMode()
{
Destroy(playerGO);
// Create population brains
agentGenomeList = new List <AgentGenome>();
BodyGenome templateBodyGenome = new BodyGenome();
templateBodyGenome.InitializeGenomeAsDefault();
//BodyGenome bodyGenomeTemplate = new BodyGenome();
//bodyGenomeTemplate.CopyBodyGenomeFromTemplate(bodyTemplate);
for (int i = 0; i < trainingPopulationSize; i++)
{
AgentGenome newGenome = new AgentGenome(i);
newGenome.InitializeBodyGenomeFromTemplate(templateBodyGenome);
newGenome.InitializeRandomBrainFromCurrentBody(0.033f);
agentGenomeList.Add(newGenome);
}
// CrossoverSettings:
trainingSettingsManager = new TrainingSettingsManager(0.06f, 0.35f, 0.1f, 0.01f);
GameObject dummyAgentGO = new GameObject("DummyAgent");
dummyAgent = dummyAgentGO.AddComponent <Agent>();
dummyStartGenome = new StartPositionGenome(Vector3.zero, Quaternion.identity);
ResetTrainingForNewGen();
//Brain brain = new Brain()
//Debug.Log(dummyAgent.testModule.);
// Create Visible Display Agent to observe behavior
/*displayStartPos = new StartPositionGenome(Vector3.zero, Quaternion.identity);
* GameObject agentGO = Instantiate(Resources.Load("PredatorPrey/PredatorPrefab")) as GameObject; ; // GameObject.CreatePrimitive(PrimitiveType.Sphere);
* //displayAgent
* agentGO.name = "Guinea Pig";
* //agentGO.transform.parent = gameObject.transform;
* agentGO.transform.localPosition = displayStartPos.agentStartPosition;
* agentGO.transform.localRotation = displayStartPos.agentStartRotation;
* //agentGO.GetComponent<Collider>().enabled = false;
* displayAgent = agentGO.AddComponent<Agent>();
* displayAgent.isVisible = true;
* displayAgent.InitializeAgentFromTemplate(agentGenomeList[0], displayStartPos);
* // hook modules:
* displayAgent.testModule.enemyTransform = targetGO.transform;
* displayAgent.testModule.bias[0] = 1f;
* displayAgent.testModule.enemyTestModule = displayAgent.testModule; // self as target, to zero out targetvel
*/
ResetTrainingAgentAndEnvironment();
}
public void SetStatusFromData(TrainingManager trainerRef)
{
this.trainerRef = trainerRef;
int focusPop = trainerRef.evaluationManager.exhibitionTicketList[trainerRef.evaluationManager.exhibitionTicketCurrentIndex].focusPopIndex;
int numPerfReps = 0;
int numHistReps = 0;
int numBaseReps = 0;
bool isEvolving;
if (focusPop < 1)
{
// env
numPerfReps = trainerRef.teamsConfig.environmentPopulation.numPerformanceReps;
numHistReps = trainerRef.teamsConfig.environmentPopulation.numHistoricalReps;
numBaseReps = trainerRef.teamsConfig.environmentPopulation.numBaselineReps;
currentTrainingSettingsRef = trainerRef.teamsConfig.environmentPopulation.trainingSettingsManager;
isEvolving = trainerRef.teamsConfig.environmentPopulation.isTraining;
}
else
{
numPerfReps = trainerRef.teamsConfig.playersList[focusPop - 1].numPerformanceReps;
numHistReps = trainerRef.teamsConfig.playersList[focusPop - 1].numHistoricalReps;
numBaseReps = trainerRef.teamsConfig.playersList[focusPop - 1].numBaselineReps;
currentTrainingSettingsRef = trainerRef.teamsConfig.playersList[focusPop - 1].trainingSettingsManager;
isEvolving = trainerRef.teamsConfig.playersList[focusPop - 1].isTraining;
}
textRepresentativesPerfValue.text = numPerfReps.ToString();
textRepresentativesHistValue.text = numHistReps.ToString();
textRepresentativesBaseValue.text = numBaseReps.ToString();
textMutationChanceValue.text = (100f * currentTrainingSettingsRef.mutationChance).ToString("F3") + "%";
textMutationStepSizeValue.text = (100f * currentTrainingSettingsRef.mutationStepSize).ToString("F1") + "%";
textNewLinkChanceValue.text = (100f * currentTrainingSettingsRef.newLinkChance).ToString("F2") + "%";
textNewNeuronChanceValue.text = (100f * currentTrainingSettingsRef.newHiddenNodeChance).ToString("F2") + "%";
if (isEvolving)
{
buttonEvolving.GetComponent <Image>().color = Color.white;
buttonEvolving.GetComponentInChildren <Text>().color = Color.black;
buttonEvolving.GetComponentInChildren <Text>().text = "EVOLVING";
}
else
{
buttonEvolving.GetComponent <Image>().color = Color.black;
buttonEvolving.GetComponentInChildren <Text>().color = Color.white;
buttonEvolving.GetComponentInChildren <Text>().text = "STATIC";
}
}
// Representative system will be expanded later - for now, just defaults to Top # of performers
public PlayerPopulation(int index, BodyGenome bodyTemplate, int numGenomes, int numPerfReps)
{
this.index = index;
// Re-Factor:
bodyGenomeTemplate = new BodyGenome();
bodyGenomeTemplate.CopyBodyGenomeFromTemplate(bodyTemplate);
//graphKing = new TheGraphKing();
popSize = numGenomes;
//this.numBaseline = numBaseline;
// Create blank AgentGenomes for the standard population
agentGenomeList = new List <AgentGenome>();
historicGenomePool = new List <AgentGenome>();
//baselineGenomePool = new List<AgentGenome>();
for (int j = 0; j < numGenomes; j++)
{
AgentGenome agentGenome = new AgentGenome(j);
agentGenome.InitializeBodyGenomeFromTemplate(bodyGenomeTemplate);
agentGenome.InitializeRandomBrainFromCurrentBody(0.2f);
agentGenomeList.Add(agentGenome);
}
//RepopulateBaselineGenomes();
//AppendBaselineGenomes();
// Representatives:
numPerformanceReps = numPerfReps;
//Debug.Log("historicGenomePool count b4: " + historicGenomePool.Count.ToString());
int numStartingHistoricalReps = 20;
for (int h = 0; h < numStartingHistoricalReps; h++)
{
historicGenomePool.Add(agentGenomeList[h]); // init
}
//Debug.Log("historicGenomePool count after: " + historicGenomePool.Count.ToString());
ResetRepresentativesList();
fitnessManager = new FitnessManager();
SetUpDefaultFitnessComponents(fitnessManager, this.index);
//fitnessManager.ResetHistoricalData();
//fitnessManager.ResetCurrentHistoricalDataLists();
fitnessManager.InitializeForNewGeneration(agentGenomeList.Count);
trainingSettingsManager = new TrainingSettingsManager(0.075f, 0.04f, 0.015f, 0.003f);
}
// Representative system will be expanded later - for now, just defaults to Top # of performers
public PlayerPopulation(Challenge.Type challengeType, BodyGenome bodyTemplate, int numGenomes, int numBaseline, int numReps)
{
bodyGenomeTemplate = new BodyGenome();
bodyGenomeTemplate.CopyBodyGenomeFromTemplate(bodyTemplate);
graphKing = new TheGraphKing();
popSize = numGenomes;
this.numBaseline = numBaseline;
// Create blank AgentGenomes for the standard population
agentGenomeList = new List <AgentGenome>();
historicGenomePool = new List <AgentGenome>();
baselineGenomePool = new List <AgentGenome>();
for (int j = 0; j < numGenomes; j++)
{
AgentGenome agentGenome = new AgentGenome(j);
agentGenome.InitializeBodyGenomeFromTemplate(bodyGenomeTemplate);
agentGenome.InitializeRandomBrainFromCurrentBody(0.0f);
agentGenomeList.Add(agentGenome);
}
RepopulateBaselineGenomes();
AppendBaselineGenomes();
// Representatives:
numPerformanceReps = numReps;
ResetRepresentativesList();
historicGenomePool.Add(agentGenomeList[0]); // init
fitnessManager = new FitnessManager();
SetUpDefaultFitnessComponents(challengeType, fitnessManager);
fitnessManager.ResetHistoricalData();
fitnessManager.ResetCurrentHistoricalDataLists();
fitnessManager.InitializeForNewGeneration(agentGenomeList.Count);
trainingSettingsManager = new TrainingSettingsManager(0.01f, 0.8f, 0.2f, 0.005f);
}
//public int numBaselineReps = 0;
public EnvironmentPopulation(EnvironmentGenome templateGenome, int numGenomes, int numReps)
{
this.templateGenome = templateGenome;
popSize = numGenomes;
//this.numBaseline = numBaseline;
environmentGenomeList = new List <EnvironmentGenome>();
historicGenomePool = new List <EnvironmentGenome>();
//baselineGenomePool = new List<EnvironmentGenome>();
for (int e = 0; e < numGenomes; e++)
{
// Create new environmentGenome
EnvironmentGenome envGenome = new EnvironmentGenome(e);
envGenome.InitializeRandomGenomeFromTemplate(templateGenome);
// Add to envGenomesList:
environmentGenomeList.Add(envGenome);
// Create parallel initial batch of genomes to be used as baseline comparison
//EnvironmentGenome baseGenome = new EnvironmentGenome(e);
//baseGenome.InitializeRandomGenomeFromTemplate(templateGenome);
//baselineGenomePool.Add(baseGenome);
}
//AppendBaselineGenomes();
// Representatives:
numPerformanceReps = numReps;
ResetRepresentativesList();
historicGenomePool.Add(environmentGenomeList[0]); // init
fitnessManager = new FitnessManager();
SetUpDefaultFitnessComponents(fitnessManager);
//fitnessManager.ResetHistoricalData();
fitnessManager.InitializeForNewGeneration(environmentGenomeList.Count);
trainingSettingsManager = new TrainingSettingsManager(0.25f, 0.05f, 0f, 0f);
}
public void SetToMutatedCopyOfParentGenome(BrainGenome parentGenome, TrainingSettingsManager settings)
{
this.bodyNeuronList = parentGenome.bodyNeuronList; // UNSUSTAINABLE!!! might work now since all neuronLists are identical
// Alternate: SetBodyNeuronsFromTemplate(BodyGenome templateBody);
// Existing Hidden Neurons!!
hiddenNeuronList = new List <NeuronGenome>();
for (int i = 0; i < parentGenome.hiddenNeuronList.Count; i++)
{
NeuronGenome newHiddenNeuronGenome = new NeuronGenome(parentGenome.hiddenNeuronList[i]); // create new neuron as a copy of parent neuron
// Might be able to simply copy hiddenNeuronList or individual hiddenNeuronGenomes from parent if they are functionally identical...
// for now going with the thorough approach of a reference-less copy
hiddenNeuronList.Add(newHiddenNeuronGenome);
}
// Existing Links!!
linkList = new List <LinkGenome>();
for (int i = 0; i < parentGenome.linkList.Count; i++)
{
LinkGenome newLinkGenome = new LinkGenome(parentGenome.linkList[i].fromModuleID, parentGenome.linkList[i].fromNeuronID, parentGenome.linkList[i].toModuleID, parentGenome.linkList[i].toNeuronID, parentGenome.linkList[i].weight, true);
float randChance = UnityEngine.Random.Range(0f, 1f);
if (randChance < settings.mutationChance)
{
float randomWeight = Gaussian.GetRandomGaussian();
newLinkGenome.weight = Mathf.Lerp(newLinkGenome.weight, randomWeight, settings.mutationStepSize);
}
linkList.Add(newLinkGenome);
}
// Add Brand New Link:
//
float randLink = UnityEngine.Random.Range(0f, 1f);
if (randLink < settings.newLinkChance)
{
List <NeuronGenome> inputNeuronList = new List <NeuronGenome>();
//List<NeuronGenome> hiddenNeuronList = new List<NeuronGenome>();
List <NeuronGenome> outputNeuronList = new List <NeuronGenome>();
for (int j = 0; j < bodyNeuronList.Count; j++)
{
if (bodyNeuronList[j].neuronType == NeuronGenome.NeuronType.In)
{
inputNeuronList.Add(bodyNeuronList[j]);
}
if (bodyNeuronList[j].neuronType == NeuronGenome.NeuronType.Out)
{
outputNeuronList.Add(bodyNeuronList[j]);
}
}
for (int j = 0; j < hiddenNeuronList.Count; j++)
{
inputNeuronList.Add(hiddenNeuronList[j]);
outputNeuronList.Add(hiddenNeuronList[j]);
}
// Try x times to find new connection -- random scattershot approach at first:
// other methods:
// -- make sure all bodyNeurons are fully-connected when modifying body
int maxChecks = 8;
for (int k = 0; k < maxChecks; k++)
{
int randID = UnityEngine.Random.Range(0, inputNeuronList.Count);
NID fromNID = inputNeuronList[randID].nid;
randID = UnityEngine.Random.Range(0, outputNeuronList.Count);
NID toNID = outputNeuronList[randID].nid;
// check if it exists:
bool linkExists = false;
for (int l = 0; l < linkList.Count; l++)
{
if (linkList[l].fromModuleID == fromNID.moduleID && linkList[l].fromNeuronID == fromNID.neuronID && linkList[l].toModuleID == toNID.moduleID && linkList[l].toNeuronID == toNID.neuronID)
{
linkExists = true;
break;
}
}
if (linkExists)
{
}
else
{
float randomWeight = Gaussian.GetRandomGaussian() * 0f;
LinkGenome linkGenome = new LinkGenome(fromNID.moduleID, fromNID.neuronID, toNID.moduleID, toNID.neuronID, randomWeight, true);
//Debug.Log("New Link! from: [" + fromNID.moduleID.ToString() + ", " + fromNID.neuronID.ToString() + "], to: [" + toNID.moduleID.ToString() + ", " + toNID.neuronID.ToString() + "]");
linkList.Add(linkGenome);
break;
}
}
}
// Add Brand New Hidden Neuron:
float randNeuronChance = UnityEngine.Random.Range(0f, 1f);
if (randNeuronChance < settings.newHiddenNodeChance)
{
// find a link and expand it:
int randLinkID = UnityEngine.Random.Range(0, linkList.Count);
// create new neuron
NeuronGenome newNeuronGenome = new NeuronGenome(NeuronGenome.NeuronType.Hid, -1, hiddenNeuronList.Count);
hiddenNeuronList.Add(newNeuronGenome);
// create 2 new links
LinkGenome linkGenome1 = new LinkGenome(linkList[randLinkID].fromModuleID, linkList[randLinkID].fromNeuronID, newNeuronGenome.nid.moduleID, newNeuronGenome.nid.neuronID, 1f, true);
LinkGenome linkGenome2 = new LinkGenome(newNeuronGenome.nid.moduleID, newNeuronGenome.nid.neuronID, linkList[randLinkID].toModuleID, linkList[randLinkID].toNeuronID, linkList[randLinkID].weight, true);
// delete old link
linkList.RemoveAt(randLinkID);
// add new links
linkList.Add(linkGenome1);
linkList.Add(linkGenome2);
//Debug.Log("New Neuron! " + newNeuronGenome.nid.neuronID.ToString() + " - from: [" + linkGenome1.fromModuleID.ToString() + ", " + linkGenome1.fromNeuronID.ToString() + "], to: [" + linkGenome2.toModuleID.ToString() + ", " + linkGenome2.toNeuronID.ToString() + "]");
}
}