private Int64 Init(RlmNetwork rnn_net, Int64 sessionID, List <RlmIOWithValue> inputs_values, Boolean learn, List <RlmIOWithValue> output_values = null)
{
if (this.CaseReference != null)
{
throw new Exception("You may only envoke Runxx once per cycle object.");
}
// set type of learning based on Inputs and Outputs
RlmType = RlmNetworkType.Predict;
if (learn)
{
if ((inputs_values != null && inputs_values.Count > 0) &&
(output_values != null && output_values.Count > 0)) // Inputs & Outputs
{
RlmType = RlmNetworkType.Supervised;
}
else if ((inputs_values != null && inputs_values.Count > 0) &&
(output_values == null || (output_values != null && output_values.Count == 0))) // Inputs without Outputs
{
RlmType = RlmNetworkType.Unsupervised;
}
}
Case cyclecase = new Case()
{
Session_ID = sessionID /*db.Sessions.Where(item=>item.ID == sessionID).First()*/, CycleStartTime = DateTime.Now, CycleEndTime = DateTime.Now
};
this.CaseReference = cyclecase;
return(cyclecase.ID);
}
private RlmCyclecompleteArgs RunCycle(RlmNetwork rnnNet, RlmCycle cycle, List <RlmIOWithValue> inputValues, RlmNetworkType rnnType, List <RlmIOWithValue> outputValues, double cycleScore, IEnumerable <RlmIdea> ideas = null)
{
//rnn_net._AddCycleToCurrentCycles(this);
RlmCycleOutput caseOutput = RlmCerebralCortex.CoreCycleProcess(rnnNet, this, inputValues, RlmType, outputValues, cycleScore, ideas);
return(rnnNet.EndCycle(caseOutput, RlmType));
}
/// <summary>
/// starts training
/// </summary>
/// <param name="rnnNet">current network being used</param>
/// <param name="sessionID">unique identifier for the session being started</param>
/// <param name="inputsValues">Inputs with stored values</param>
/// <param name="learn">Indicator that if true, will start training, if false, will run prediction</param>
/// <param name="outputValues">Outputs with stored values</param>
/// <param name="cyclescore">Score of the current cycle</param>
/// <param name="Parallel"></param>
/// <param name="ideas">Gives bias to the RLM on what to output</param>
/// <returns></returns>
public RlmCyclecompleteArgs RunCycle(RlmNetwork rnnNet, Int64 sessionID, List <RlmIOWithValue> inputsValues, Boolean learn, List <RlmIOWithValue> outputValues = null, double cyclescore = 0.000, Boolean Parallel = false, IEnumerable <RlmIdea> ideas = null)
{
//Run Validity Checks -- includes local var assignment
Int64 cyclecaseid = Init(rnnNet, sessionID, inputsValues, learn, outputValues);
//if parallel, it is executed on another thread and returns null since the cycle complete arguments can only be taken via the Rnetwork cycle complete event
if (Parallel)
{
cycletask = Task.Run(() => RunCycle(rnnNet, this, inputsValues, RlmType, outputValues, cyclescore, ideas));
return(null);
}
else //if not parallel, we run it on same thread and return the cycle complete arguments **did it this way for a better windowless training
{
return(RunCycle(rnnNet, this, inputsValues, RlmType, outputValues, cyclescore, ideas));
}
#region Old code commented
//cycletask = Task.Run(() =>
//{
// rnn_net._AddCycleToCurrentCycles(this);
// rlm_cycle_output caseOutput = rnn_cerebralcortex.CoreCycleProcess(rnn_net, this, inputs_values, RlmType, output_values, cyclescore);
// rnn_net._EndCycle(caseOutput, RlmType);
//});
//Non-Parallel Waits for Thread to Complete
//if (!Parallel) cycletask.Wait();
#endregion
}
private void ClearNetwork()
{
if (network != null)
{
network.DataPersistenceComplete -= Network_DataPersistenceComplete;
network.Dispose();
network = null;
}
}
internal static RlmCycleOutput CoreCycleProcess(RlmNetwork rnn_net, RlmCycle rnn_cyc, IEnumerable <Models.RlmIOWithValue> rnn_ins, RlmNetworkType rnnType, IEnumerable <Models.RlmIOWithValue> rnn_outs, double cyclescore, IEnumerable <RlmIdea> ideas = null, IEnumerable <long> excludeSolutions = null)
{
var memoryMgr = rnn_net.MemoryManager;
// temp benchmark only
//rnn_net.CurrentCycleCount++;
// temp benhcmark only
// Determine if any inputs are of Linear type
bool hasLinearInputs = rnn_ins.Any(a => a.Type == RlmInputType.Linear);
// update input momentums
if (hasLinearInputs)
{
foreach (var item in rnn_ins)
{
if (item.Type == RlmInputType.Linear)
{
var inputMomentumObj = rnn_net.InputMomentums[item.ID];
inputMomentumObj.SetInputValue(Convert.ToDouble(item.Value));
item.InputMomentum = inputMomentumObj;
}
}
}
//Get rneuron
GetRneuronResult rneuronFound = memoryMgr.GetRneuronFromInputs(rnn_ins, rnn_net.CurrentNetworkID);
Rneuron neuron = rneuronFound.Rneuron;
//Holds the solution instance
GetSolutionResult solutionFound = new GetSolutionResult();
Solution solution = null;
IEnumerable <Models.RlmIO> outputs = rnn_net.Outputs;
bool completelyRandom = false;
double randomnessValue = rnn_net.RandomnessCurrentValue;
if (rnnType == RlmNetworkType.Supervised)
{
//Supervised, get solution and record ideal score
solutionFound = memoryMgr.GetSolutionFromOutputs(rnn_outs);
solution = solutionFound.Solution;
cyclescore = IDEAL_SCORE;
}
else if (rnnType == RlmNetworkType.Unsupervised && randomnessValue > 0)
{
//TODO: This should be based upon the randomization factor
double randomProbability = Util.GetRandomDoubleNumber(0, 100);
bool random = randomProbability <= randomnessValue;
//Idea
//ToDo: Implement Ideas
//The idea implementation will not be added until core functionality works. It is an "extra" and the network can learn without it. In fact, since it reduces load, we need
//to test without it in place first. Otherwise networks that don't have an applicable "idea" may crash
//System.Diagnostics.Debug.WriteLine("Threshold: " + randomnThreshold);
long?bestSolutionId = null;
if (!random)
{
// get best solution
solution = memoryMgr.GetBestSolution(rnn_ins, (hasLinearInputs) ? rnn_net.LinearToleranceCurrentValue : 0, excludeSolutions: excludeSolutions); //db.GetBestSolution(rnn_net.CurrentNetworkID, rnn_ins, (hasLinearInputs) ? rnn_net.LinearToleranceCurrentValue : 0);
bestSolutionId = solution?.ID;
if (solution == null)
{
completelyRandom = true;
solutionFound = memoryMgr.GetRandomSolutionFromOutput(randomnessValue, outputs, bestSolutionId, ideas);
}
else
{
solutionFound.Solution = solution;
solutionFound.ExistsInCache = true;
}
}
else if (random && outputs.Count() > 1)
{
solution = memoryMgr.GetBestSolution(rnn_ins, (hasLinearInputs) ? rnn_net.LinearToleranceCurrentValue : 0, excludeSolutions: excludeSolutions); //db.GetBestSolution(rnn_net.CurrentNetworkID, rnn_ins, (hasLinearInputs) ? rnn_net.LinearToleranceCurrentValue : 0);
bestSolutionId = solution?.ID;
completelyRandom = true;
solutionFound = memoryMgr.GetRandomSolutionFromOutput(randomnessValue, outputs, bestSolutionId, ideas);
}
else
{
completelyRandom = true;
solutionFound = memoryMgr.GetRandomSolutionFromOutput(randomnessValue, outputs, ideas: ideas);
}
solution = solutionFound.Solution;
}
else // Predict
{
solution = memoryMgr.GetBestSolution(rnn_ins, predict: true, predictLinearTolerance: rnn_net.PredictLinear, excludeSolutions: excludeSolutions); //db.GetBestSolution(rnn_net.CurrentNetworkID, new List<long>() { neuron.ID }, true);
if (solution == null)
{
completelyRandom = true;
solutionFound = memoryMgr.GetRandomSolutionFromOutput(randomnessValue, outputs, ideas: ideas);
solution = solutionFound.Solution;
#region TODO cousin node search
//// no solution found AND all inputs are Distinct
//if (!hasLinearInputs)
//{
// completelyRandom = true;
// //solution = GetRandomSolutionFromOutput(db, rnn_net.CurrentNetworkID, outputs, false);
// solutionFound = memoryMgr.GetRandomSolutionFromOutput(randomnessValue, outputs); //GetRandomSolutionFromOutput(db, rnn_net, outputs, rnn_ins, (hasLinearInputs) ? rnn_net.LinearToleranceCurrentValue : 0);
//}
//else // has linear
//{
// // TODO need to change the methods used below to MemoryManager
// //// gets all the known inputs
// //var knownInputs = DetermineKnownInputs(db, rnn_ins, rnn_net.CousinNodeSearchToleranceIncrement);
// //if (knownInputs.Count > 0)
// //{
// // // holds the top cases for each known input
// // var topCases = new List<Case>();
// // foreach (var item in knownInputs)
// // {
// // // gets the top solution for the current input with incremental checks based on the linear bracket
// // var topCase = GetBestKnownCase(db, item, rnn_net.CousinNodeSearchToleranceIncrement);
// // if (topCase != null)
// // {
// // topCases.Add(topCase);
// // }
// // }
// // // determine which Case has the highest score and get it's corresponding solution
// // solution = topCases.OrderByDescending(a => a.Session.DateTimeStop)
// // .ThenByDescending(a => a.CycleEndTime)
// // .ThenByDescending(a => a.CycleScore)
// // .ThenByDescending(a => a.Session.SessionScore)
// // .Take(1)
// // .Select(a => a.Solution)
// // .FirstOrDefault();
// //}
// //else // if no known inputs then we get solution randomly
// //{
// // completelyRandom = true;
// // //solution = GetRandomSolutionFromOutput(db, rnn_net.CurrentNetworkID, outputs, false);
// // solution = GetRandomSolutionFromOutput(db, rnn_net, outputs, rnn_ins, (hasLinearInputs) ? rnn_net.LinearToleranceCurrentValue : 0);
// //}
//}
#endregion
//solutionFound.Solution = solution;
//solutionFound.ExistsInCache = false;
}
else
{
solutionFound.Solution = solution;
solutionFound.ExistsInCache = true;
}
}
//Document score, solution in Case
var newCase = RecordCase(rnn_cyc
, rneuronFound
, rnn_ins
, rnn_outs
, cyclescore
, solutionFound
, 0 //ToDo: Pass the current maturity factor setting
, completelyRandom //ToDo: pass whether or not the result was completely randomly generated
, 0 //ToDo: pass sequential count
);
// set Current case reference
rnn_net.CurrentCase = newCase;
var cycleOutput = new RlmCycleOutput(newCase.ID, newCase.Rneuron_ID, newCase.Solution_ID, rnn_net.Outputs, solution.Output_Values_Solutions);
cycleOutput.CompletelyRandom = completelyRandom;
return(cycleOutput);
}
private IDictionary <int, RlmCyclecompleteArgs> RunOneSession(RlmNetwork network, IEnumerable <KeyValuePair <string, Resource> > resourceInputs, bool learn = true, bool showCycleOutput = false)
{
CycleOutputs.Clear();
SessionOutputs.Clear();
IDictionary <int, RlmCyclecompleteArgs> cycleOutputDic = new Dictionary <int, RlmCyclecompleteArgs>();
TrainingVariables["SessionScore"].Value = 0;
long sessId = network.SessionStart();
List <object> outputList = new List <object>();
//TODO: check later for multiple inputs
int min = 0;
int max = 0;
var resIn = resourceInputs.First();
Resource res = resIn.Value;
Dictionary <string, int> inputRange = GetInputRange(res);
min = inputRange["Min"];
max = inputRange["Max"];
bool usedExcludeSolutions = false;
for (int j = min; j <= max; j++)
{
RlmCyclecompleteArgs result;
List <long> excludeSolutions = null;
while (true)
{
//Populate input values
var invs = new List <RlmIOWithValue>();
foreach (var a in network.Inputs)
{
invs.Add(new RlmIOWithValue(a, j.ToString()));
CycleInputs[a.Name] = j;
}
//Build and run a new RlmCycle
var Cycle = new RlmCycle();
result = Cycle.RunCycle(network, sessId, invs, learn, excludeSolutions: excludeSolutions);
//TODO: check later for multiple outputs
var rlmOut = network.Outputs.First();
var value = result.CycleOutput.Outputs.First(b => b.Name == rlmOut.Name).Value;
if (!AllowDuplicates && !learn)
{
bool hasDup = false;
foreach (var rlmOutputs in SessionOutputs)
{
if (rlmOutputs.Value.Any(a => a.ToString() == value))
{
if (excludeSolutions == null)
{
excludeSolutions = new List <long>();
}
excludeSolutions.Add(result.CycleOutput.SolutionID);
hasDup = true;
System.Diagnostics.Debug.WriteLine("Duplicate found!");
break;
}
}
if (hasDup)
{
usedExcludeSolutions = true;
continue;
}
}
//set current rn and sl
cycleOutputDic[j] = result;
CycleOutputs[rlmOut.Name] = value;
outputList.Add(value);
SessionOutputs[rlmOut.Name] = outputList;
break;
}
double cycleScore = ScoreCycle();
if (showCycleOutput)
{
//if (j == min && PredictData.Count() > min)
//{
// PredictData.Clear();
//}
//// exposed cyclescore on predict
//PredictData.Add(cycleScore.ToString());
}
network.ScoreCycle(result.CycleOutput.CycleID, cycleScore);
TrainingVariables["CycleScore"].Value = 0;
}
double sessionScore = ScoreSession();
LastScore = sessionScore;
if (LastScore > HighScore)
{
HighScore = LastScore;
}
network.SessionEnd(sessionScore);
TraceLogObjectsIf(!learn && !usedExcludeSolutions && sessionScore < HighScore, $"{network.DatabaseName}_BestSolutions_InMemory", network.MemoryManager.BestSolutions.SelectMany(a => a.Value.Values));
TraceLog($"Score: {string.Format("{0:n}", sessionScore)}");
// exposed session data
//SessionData.Add(sessionScore.ToString());
return(cycleOutputDic);
}
public IDictionary <int, RlmCyclecompleteArgs> StartTraining(RlmSettings settings, CancellationToken?token = null)
{
dataPersistEvent.Reset();
//stopTraining = false;
//settings = this.Settings;
this.Settings = settings;
//Get RLM Settings from user
int startRandomness = settings.StartRandomness;
int endRandomness = settings.EndRandomness;
int maxBracket = settings.MaxLinearBracket;
int minBracket = settings.MinLinearBracket;
RlmSimulationType simType = settings.SimulationType;
TimeSpan time = settings.Time;
double target = settings.SimulationTarget;
// clear previous network, if any
ClearNetwork();
IDictionary <int, RlmCyclecompleteArgs> cycleOutputDic = null;
network = (string.IsNullOrEmpty(DatabaseName) ? new RlmNetwork() : new RlmNetwork(RlmDbData, true));
network.DataPersistenceComplete += Network_DataPersistenceComplete;
network.StartRandomness = startRandomness;
network.EndRandomness = endRandomness;
network.MaxLinearBracket = maxBracket;
network.MinLinearBracket = minBracket;
List <RlmIO> inputs = new List <RlmIO>();
List <RlmIO> outputs = new List <RlmIO>();
//get inputs/outputs from resources
var resourceInputs = Resources.Where(a => a.Value.RLMObject == RLMObject.Input);
var resourceOutputs = Resources.Where(a => a.Value.RLMObject == RLMObject.Output);
//set rlm inputs/outputs
inputs = GetRlmIOFromResource(resourceInputs);
outputs = GetRlmIOFromResource(resourceOutputs);
//create network
if (!network.LoadNetwork())
{
network.NewNetwork("RlmOptimizer", inputs, outputs);
}
//start training
int sessions = settings.SimulationType == RlmSimulationType.Sessions ? Convert.ToInt32(settings.SimulationTarget) : settings.NumOfSessionsReset;
double hours = settings.SimulationType == RlmSimulationType.Time ? settings.Time.TotalHours : 0;
DateTime endsOn = DateTime.Now.AddHours(hours);
TraceLog($"RLM Settings: \n\n" +
$"Randomness: Max = {startRandomness}, End = {endRandomness}\nLinear Bracket: Max = {maxBracket}, Min = {minBracket}\n" +
$"Simulation Type: {simType}\nTarget: {target}\nTime: {hours}\n" +
$"Cycle Score Formula: {string.Join(" | ", CyclePhase.Formula)}\n" +
$"Session Score Formula: {string.Join(" | ", SessionPhase.Formula)}");
TraceLog();
int scoreHits = 0;
bool timesUp = false;
//for (var trial = 1; trial <= 5; trial++)
//{
//TrainingLog($"\nTrial #{trial}");
do
{
//settings
network.NumSessions = sessions;
network.StartRandomness = startRandomness;
network.EndRandomness = endRandomness;
network.MaxLinearBracket = maxBracket;
network.MinLinearBracket = minBracket;
network.ResetRandomizationCounter();
if ((simType == RlmSimulationType.Time && DateTime.Compare(DateTime.Now, endsOn) == 1) || token?.IsCancellationRequested == true)
{
timesUp = true;
}
if (!timesUp)
{
TraceLog("\nTraining...\n");
for (int i = 1; i <= sessions; i++)
{
cycleOutputDic = RunOneSession(network, resourceInputs, true);
if (settings.SimulationType == RlmSimulationType.Score)
{
if (UpdateScoreHit(ref scoreHits, settings) || token?.IsCancellationRequested == true)
{
break;
}
}
if ((simType == RlmSimulationType.Time && DateTime.Compare(DateTime.Now, endsOn) == 1) || token?.IsCancellationRequested == true)
{
timesUp = true;
break;
}
if (stopTraining || token?.IsCancellationRequested == true)
{
break;
}
}
}
//predict
TraceLog("\nPredicting...\n");
if (settings.SimulationType == RlmSimulationType.Score)
{
for (int i = 1; i <= settings.NumScoreHits; i++)
{
cycleOutputDic = RunOneSession(network, resourceInputs, false, (i == settings.NumScoreHits));
if (UpdateScoreHit(ref scoreHits, settings) || token?.IsCancellationRequested == true)
{
break;
}
}
}
else
{
cycleOutputDic = RunOneSession(network, resourceInputs, false, true);
}
if (stopTraining || token?.IsCancellationRequested == true)
{
break;
}
} while ((!timesUp && simType == RlmSimulationType.Time) ||
(simType == RlmSimulationType.Score && settings.NumScoreHits > scoreHits));
//}
//end training
network.TrainingDone();
if (simType == RlmSimulationType.Score)
{
TraceLog($"Score Hits: {scoreHits}");
}
TraceLog("Optimization Done");
return(cycleOutputDic);
}
