public void StartRERA(DaneshWindow gen, int num)
{
danesh = gen;
numberOfAttemptsRandom = num;
// DrawRandomERGraph();
LastERA = SampleExpressiveRangeRandomly(num, gen);
}
public float Evaluate(object[] arr, float samples, DaneshWindow dan)
{
for (int i = 0; i < c_params.Count; i++)
{
c_params[i].field.SetValue(c_params[i].owner, arr[i]);
}
float totalScore = 0f;
for (int att = 0; att < samples; att++)
{
object map = dan.GenerateContent();
float score = 0f;
//Obtain a metric reading for every targeted metric
for (int i = 0; i < c_metrics.Count; i++)
{
float val = (float)dan.GetMetric(c_metrics[i], new object[] { map });
score += 1 - Mathf.Abs(val - c_targets[i]);
}
//Get a score by averaging them
totalScore += score / c_metrics.Count;
}
return(totalScore / (float)samples);
}
List <List <float> > SampleExpressiveRangeRandomly(int totalAttempts, DaneshWindow gen)
{
float progressBar = 0f;
EditorUtility.DisplayProgressBar("Computing Randomised Expressive Range Histogram", "Working...", progressBar);
List <List <float> > res = new List <List <float> >();
/*
* Current:
* eraSample[metric1][metric2] = object[,] of samples representing the histogram
* Lot of data duplication here, not the most efficient way to do it
*/
eraSamples = new List <List <object[, ]> >();
for (int i = 0; i < danesh.metricList.Count; i++)
{
List <object[, ]> secondmetric = new List <object[, ]>();
for (int j = 0; j < danesh.metricList.Count; j++)
{
object[,] sampleH = new object[100, 100];
secondmetric.Add(sampleH);
}
eraSamples.Add(secondmetric);
}
for (int att = 0; att < totalAttempts; att++)
{
//Randomly parameterise the generator
foreach (GeneratorParameter p in gen.parameterList)
{
p.RandomiseValue();
}
object map = danesh.GenerateContent();
List <float> nums = new List <float>();
for (int i = 0; i < danesh.metricList.Count; i++)
{
float score = (float)danesh.GetMetric(i, new object[] { map });
nums.Add(score);
}
//Update the samples list
for (int i = 0; i < nums.Count; i++)
{
int index1 = (int)Mathf.Round(nums[i] * 100f);
if (index1 < 0)
{
index1 = 0;
}
if (index1 > 99)
{
index1 = 99;
}
for (int j = 0; j < nums.Count; j++)
{
int index2 = (int)Mathf.Round(nums[j] * 100f);
if (index2 < 0)
{
index2 = 0;
}
if (index2 > 99)
{
index2 = 99;
}
eraSamples[i][j][index1, index2] = map;
}
}
res.Add(nums);
EditorUtility.DisplayProgressBar("Computing Randomised Expressive Range Histogram", "Evaluating random expressive range... " + (100 * (float)att / (float)totalAttempts).ToString("F0") + " percent complete", (float)att / (float)totalAttempts);
}
EditorUtility.ClearProgressBar();
return(res);
}
public void StartERA(DaneshWindow gen, int num)
{
danesh = gen;
numberOfAttempts = num;
LastERA = SampleExpressiveRange(num);
}
public List <GeneratorParameter> FindParameters(DaneshWindow d, MonoBehaviour b)
{
// Debug.Log("Searching for parameters... "+d.metricList.Count+" metrics found.");
List <GeneratorParameter> res = new List <GeneratorParameter>();
//Collect a control sample set
List <float> metricAverages = new List <float>();
List <float[]> stdev_samples = new List <float[]>();
List <float> stdev = new List <float>();
for (int i = 0; i < d.metricList.Count; i++)
{
metricAverages.Add(0);
stdev_samples.Add(new float[numSamples]);
}
for (int s = 0; s < numSamples; s++)
{
object output = d.GenerateContent();
for (int i = 0; i < d.metricList.Count; i++)
{
GeneratorMetric m = d.metricList[i];
float sc = (float)m.method.Invoke(null, new object[] { output });
metricAverages[i] += sc;
stdev_samples[i][s] = sc;
}
}
//Calculate averages and standard deviation
for (int i = 0; i < d.metricList.Count; i++)
{
metricAverages[i] = metricAverages[i] / numSamples;
// Debug.Log("Average for metric "+i+": "+metricAverages[i]);
float sqdif = 0;
foreach (float s in stdev_samples[i])
{
sqdif += Mathf.Pow(s - metricAverages[i], 2f);
}
sqdif = sqdif / numSamples;
stdev.Add(Mathf.Sqrt(sqdif));
// Debug.Log("Standard deviation for metric "+i+": "+stdev[i]);
}
foreach (FieldInfo field in d.generator.GetType().GetFields())
{
bool useParam = true;
foreach (Attribute _attr in field.GetCustomAttributes(false))
{
if (_attr is TunableAttribute)
{
//We already know this parameter is good so don't worry.
useParam = false;
}
}
if (!useParam)
{
continue;
}
//Remember the original value
object original_value = field.GetValue(b);
//Change the value to something
List <object> sampleValues = GetSampleValues(field, b);
// Debug.Log(sampleValues.Count+" sample values found");
//For each sample value, change the field to that value, run samples, get average
foreach (object o in sampleValues)
{
field.SetValue(b, o);
List <float> sampleData = new List <float>();
for (int i = 0; i < d.metricList.Count; i++)
{
sampleData.Add(0);
}
bool sampleFailed = false;
for (int _ = 0; _ < numSamples; _++)
{
try{
object output = d.GenerateContent();
for (int i = 0; i < d.metricList.Count; i++)
{
GeneratorMetric m = d.metricList[i];
sampleData[i] += (float)m.method.Invoke(null, new object[] { output });
}
}
catch (Exception e) {
//This failed, check the next value
sampleFailed = true;
break;
}
}
if (sampleFailed)
{
continue;
}
//Did this change any of the metrics more than one standard deviation from the original sampling?
bool nextParam = false;
for (int i = 0; i < d.metricList.Count; i++)
{
sampleData[i] = sampleData[i] / numSamples;
if (Mathf.Abs(sampleData[i] - metricAverages[i]) > stdev[i])
{
Debug.Log("Parameter " + field.Name + " may have an impact on metrics");
nextParam = true;
GeneratorParameter p = new GeneratorParameter(field.Name, original_value, original_value, original_value, field, b);
p.locked = true;
res.Add(p);
break;
}
}
if (nextParam)
{
field.SetValue(b, original_value);
break;
}
}
field.SetValue(b, original_value);
}
return(res);
}
public void Tune(DaneshWindow dan, List <GeneratorParameter> ps, List <GeneratorMetric> ms, List <float> ts)
{
this.c_params = ps;
this.c_metrics = ms;
this.c_targets = ts;
float bestScore = 0f;
object[] bestArray = new object[ps.Count];
//Pick a random spot
float time = Time.realtimeSinceStartup;
object[] ex = new object[ps.Count];
for (int s = 0; s < ps.Count; s++)
{
ex[s] = ps[s].GetRandomValue();
}
bestScore = Evaluate(ex, 50, dan);
bestArray = ex;
EditorUtility.DisplayProgressBar("Automatic Tuning In Progress", "Working...", 0f);
for (int iter = 0; iter < iterations; iter++)
{
if (bestScore > 0.98f)
{
break;
}
//Find all of the neighbouring parameter sets
List <object[]> nbs = CalculateNeighbours(ex);
float bestThisRound = bestScore;
object[] bestArrayThisRound = new object[c_params.Count];
foreach (object[] nb in nbs)
{
float score = Evaluate(nb, 50, dan);
if (score > bestThisRound)
{
bestThisRound = score;
bestArrayThisRound = nb;
}
}
if (bestThisRound > bestScore)
{
bestScore = bestThisRound;
bestArray = bestArrayThisRound;
ex = bestArrayThisRound;
}
else
{
//Random restart
ex = new object[c_params.Count];
for (int s = 0; s < c_params.Count; s++)
{
ex[s] = ps[s].GetRandomValue();
}
}
EditorUtility.DisplayProgressBar("Auto-tuning", "Searching parameter space... " + (100 * (float)iter / (float)iterations).ToString("F0") + " percent complete", (float)iter / (float)iterations);
}
EditorUtility.ClearProgressBar();
//Apply the parameters
object[] pvs = bestArray;
for (int i = 0; i < pvs.Length; i++)
{
c_params[i].SetValue(pvs[i]);
}
}
