Skip to content

Latest commit

 

History

History
204 lines (166 loc) · 16 KB

README.md

File metadata and controls

204 lines (166 loc) · 16 KB

STGP-Sharp

STGP-Sharp is a strongly-typed genetic programming framework for C#. It is directly compatible with Unity.

STGP-Sharp is a standalone version of the genetic programming framework used in ABL-Unity3D. Currently, it provides more capabilities such as co-evolution and slightly different implementation details.

Requirements

  • DotNet 6.0

Sample Experiments

We provide sample experiments in the file SampleExperiments.cs. The experiment provided is one in which we use GP to generate a genome using the logical operators AND and NOT to create an equivalent operator to the logical operator OR.

GP Search Algorithm

We provide a standard GP search algorithm with the capabilities for co-evolution, as well as asynchronous individual evaluation. Whether the search algorithm uses co-evolution or asynchronous individual evaluation is determined by the user-provided fitness function. See here for more information on how to define a fitness function.

The GP uses reflection to find user-defined GP operators, which are then used in the search algorithm to generate individuals.

In addition, the user must provide a probability distribution for the different GP operators. This probability distribution is used by the GP search algorithm to randomly generate individuals.

Lastly, STGP-Sharp provides the option to use multiple different population initialization methods. By default, random and ramped initialization are included. The user can create their own.

NOTE: We use the terms "genome", "node", and "tree" interchangebly.

Defining a GP Operator

GpBuildingBlock<T> is the super-class for GP primitives where T is the evaluation type. GpBuildingBlock<T> extends the class Node, which is a generic class for a genome with no strong typing. GpBuildingBlock<T>.children is the list of child nodes. Child nodes extend GpBuildingBlock<T>. A GP primitive must define a constructor which passes all child nodes to the primitive base constructor. A GP primitive must also override the function Evaluate, which returns the result of the GP primitive upon evaluation. The return type of Evaluate must be the same as the evaluation type T for the GP primitive.

The following figure shows sample code for defining the Not GP primitive shown in Figure 2, which models the boolean operator NOT. The GP primitive Not has an evaluation type of bool, as shown by line 1. Line 3 is a helper properties. Lines 5-7 define the constructor for the Not class. Lastly, lines 9-12 define the Evaluate method. Line 11 evaluates the child node, and then returns the negated value.

1.  public class Not : GpBuildingBlock<bool>  
2.  {  
3.      public GpBuildingBlock<bool> Operand => (GpBuildingBlock<bool>)this.children[0]; 
4. 
5.      public Not(GpBuildingBlock<bool> operand) : base(operand)  
6.      {  
7.      }  
8. 
9.      public override bool Evaluate(GpFieldsWrapper gpFieldsWrapper)  
10.     {
11.          return !this.Operand.Evaluate(gpFieldsWrapper); 
12.     }
13.  }

Child nodes of a GP primitive can be immutable; genetic operators can not modify them. For example, replace line 6 with a hard-coded instance of a subclass of GpBuildingBlock<bool>. E.g., an instance of the GP primitive BooleanConstant.

Fitness

There are different types of fitness values. For example, the following project implements lexicographic fitness.

Each fitness type must extend the class FitnessBase.

Given that there are different types of fitness values, it may or may not make sense for certain types of fitness values to be compared to other fitness types. Thus, the FitnessBase class provides a helper function ThrowExceptionIfInvalidFitnessForComparison which throws an exception if two incompatible types of fitness are compared.

Lastly, each fitness type also has an associated type to generate statistics for that fitness type. For example, lexicographic fitness can define summary statistics for each dimension of the fitness value. To do so, you must define the static property GpResultsStatsType (see here for my details).

Below is a partial implementation of the FitnessStandard class. We only show part of the implementation so as to illustrate the most important pieces. Line 1 defines the class by extending FitnessBase. Lines 3-8 define the fitness value score. Line 10 defines the associated GpResultsStats class for the FitnessStandard class. Lines 12-15 define a helper function to determine if the a fitness type is of type FitnessStandard. Otherwise, it throws an exception. Line 17-26 define the CompareTo function which is called when comparing two fitness values. Lines 19-22 check if the fitness to compared against is null. Line 24 calls the helper function ThrowExceptionIfInvalidFitness to verify the fitness being compared is of the correct type (e.g. FitnessStandard). Lastly, line 25 compares the fitness score of the current FitnessStandard instance to the fitness value being compared to.

1.  public class FitnessStandard : FitnessBase  
2.  {  
3.       public readonly double fitnessScore;  
4.   
5.       public FitnessStandard(double fitnessScore)  
6.       {  
7.           this.fitnessScore = fitnessScore;  
8.       }  
9.   
10.      public new static Type GpResultsStatsType { get; } = typeof(GpResultsStatsStandard);  
11. 
12.      public static void ThrowExceptionIfInvalidFitness(FitnessBase? f, out FitnessStandard fs)  
13.      {  
14.          ThrowExceptionIfInvalidFitnessForComparison(f, out fs);  
15.      }  
16.   
17.      public override int CompareTo(FitnessBase? other)  
18.      {  
19.          if (null == other)  
20.          {  
21.              return this.fitnessScore.CompareTo(null);  
22.          }  
23. 
24.          ThrowExceptionIfInvalidFitness(other, out FitnessStandard otherStandard);  
25.          return this.fitnessScore.CompareTo(otherStandard.fitnessScore);  
26.      }
27. 
28.      ...
29. }

Defining a Fitness Function

To define a fitness function, define a class that implements the interfaceIFitnessFunction. In addition, the sub-class must implement several different interfaces. If the fitness function works with co-evolution, it must either implement ICoevolutionAsync or ICoevolutionSync depending on whether the fitness function is asynchronous. Similarly, if it does not work with coevolution, it must implement IAsync or ISync depending on whether the fitness function is asynchronous.

If the sub-class works with co-evolution, it must implement the evaluation function GetFitnessOfPopulationUsingCoevolution/GetFitnessOfPopulationUsingCoevolutionAsync.

If the sub-class does not work with co-evolution, it must implement the evaluation function GetFitnessOfIndividual/GetFitnessOfIndividualAsync.

Additionally, each fitness function must define an associated Fitness type (see here for more details on Fitness types). To do so, define the property FitnessType.

Below is an implementation of the fitness function to model the logical OR operator using the logical operators AND and NOT. Lines 1-3 define that the fitness function is to be run synchronously. Line 3 defines that the fitness function works with fitness type FitnessStandard. Lines 6-8 check if the given individual has the correct return type, aka boolean. Essentially, the function checks the given individual against the truth table of the operator OR. Line 16 defines a PositionalArguments which is used to pass the booleans b1 and b2 to the evaluation function of the genome. Line 17 defines a wrapper which is used by the evaluation function of the genome. Lines 11 and 18-25 are used to evaluate the accuracy of the genome in modelling the logical OR operator. Line 26 then returns the result as an instance of the FitnessStandard class.

1. public class EquivalentToOr : IFitnessFunction, ISync  
2. {  
3.      public Type FitnessType { get; } = typeof(FitnessStandard);  
4.      public FitnessBase GetFitnessOfIndividual(GpRunner gp, Individual i)  
5.      {  
6.          if (i.genome is not GpBuildingBlock<bool> evaluateBoolean)  
7.          {  
8.              throw new Exception("Genome does not evaluate to boolean");  
9.          }  
10.  
11.         float fitnessScoreSoFar = 0;  
12.         foreach (bool b1 in new[] { true, false })  
13.         {  
14.             foreach (bool b2 in new[] { true, false })  
15.             {  
16.                 var positionalArguments = new PositionalArguments(b1, b2);
17.                 var gpFieldsWrapper = new GpFieldsWrapper(gp, positionalArguments: positionalArguments);  
18.                 if ((b1 || b2) == evaluateBoolean.Evaluate(gpFieldsWrapper))  
19.                 {  
20.                     fitnessScoreSoFar++;  
21.                 }  
22.             }  
23.         }  
24.  
25.         fitnessScoreSoFar /= 4;  
26.         return new FitnessStandard(fitnessScoreSoFar);  
27.     }  
28. }

GpResultsStats

Each fitness type has an associated type to generate summary statistics for it. To define a GpResultsStats classs, you must create a class which extends the GpResultsStatsBase. The class must implement the method GetDetailedSummary which returns a statistical summary represented by an interface IDetailedSummary.

Below is an implementation of the GpResultsStatsStandard class which is associated with the FitnessStandard class. Lines 1-5 define a constructor where you can pass in an enumerable of FitnessStandard instances. Lines 7-11 define a more general constructor which takes in an enumerable of FitnessBase instances. Line 8 attempts to cast the fitness values to the class FitnessStandard, and if it can't it throws an exception. If it is successful, it is passed to the base constructor defined in GpResultsStatsBase. It is necessary to define a more general constructor that dynamically checks if the input fitness values are of the correct type. This is a consequence of C#8 not supporting contra/covariant types. I would use a more recent version of C#, but I want this library to be compatible with Unity. In addition, this is a consequence of the implementation specifics of the GpResultsStatsBase class. This will ideally be changed in a future version of STGP-Sharp.

Lines 13-18 defines the method GetDetailedSummary which returns an instance of the struct DetailedSummaryStandard, defined in lines 19-34. DetailedSummaryStandard creates a fitness summary using the SimpleStats class which generates basic summary statistics.

1.  public class GpResultsStatsStandard : GpResultsStatsBase  
2.  {  
3.      public GpResultsStatsStandard(IEnumerable<FitnessStandard>? fitnessValues) : base(fitnessValues)  
4.      {  
5.      }  
6.   
7.      public GpResultsStatsStandard(IEnumerable<FitnessBase> fitnessValues) : base(  
8.           fitnessValues?.Cast<FitnessStandard>()  
9.           ?? throw new Exception($"The fitness values passed in are not of type {typeof(GpResultsStatsStandard)}"))  
10.     {  
11.     }  
12.   
13.     public override IDetailedSummary GetDetailedSummary()  
14.     {  
15.         return new DetailedSummaryStandard(this.fitnessValues?.Cast<FitnessStandard>() ??  
17.                                            throw new NullReferenceException(nameof(this.fitnessValues)));  
18.     }  
  
19.     public struct DetailedSummaryStandard : IDetailedSummary  
20.     {  
21.          public SimpleStats.Summary fitnessScoreSummary;  
22.    
23.          public DetailedSummaryStandard(IEnumerable<FitnessStandard> fitnessValues)  
24.          {  
25.              this.fitnessScoreSummary = new SimpleStats(  
26.                      fitnessValues.Select(f => f.fitnessScore))  
27.                  .GetSummary();  
28.          }  
29.  
30.          public override string ToString()  
31.          {  
32.              return $"Total Fitness Summary: {this.fitnessScoreSummary.ToString()}";  
33.          }  
34.     }  
35.  }

Defining a Population Initialization Method

To define a custom population initialization method, you must create a class which extends PopulationInitializationMethod and implements the method GetPopulation<T> where T either the type of the genome to be generated, or is the solution return type of the genomes to be generated. The method takes a GpRunner instance which implements the GP search algorithm, as well a TimeoutInfo instance which defines the method which keeps track of whether a breakout has been requested, or whether the given time limit has been reached. Because GetPopulation<T> provides asynchronous capabilities, it returns an awaitable list of individuals, ie. Task<List<Individual>>. See here for information on asynchronous operations through the use of Task.

Below is the implementation of randomized population initialization provided by STGP-Sharp. Line 5 defines the initially empty population. Line 6 create a loop which ends either when the size of population reaches the user-provided parameter populationSize, or the user-provided time limit has been reached. Then line 8-9, using the method GpRunner.GenerateRandomTreeFromTypeOrReturnType<T>, a random genome is generated of either type T or of return type T. This method takes in the maximum depth parameter provided by the user. GpRunner.GenerateRandomTreeFromTypeOrReturnType<T> takes the parameter forceFullyGrow which defines whether the random genome generated must be as large as possible. Line 10 then takes the random genome generated and creates an individual from it. Line 11 then evaluates the individual and assigns it a fitness score using the method GpRunner.EvaluateFitnessOfIndividual. Line 12 then adds this new individual to the population. Lastly, line 14 returns the generated population.

1.  public class RandomPopulationInitialization : PopulationInitializationMethod
2.  {  
3.      public override async Task<List<Individual>> GetPopulation<T>(GpRunner gp, TimeoutInfo timeoutInfo)
4.      {   
5.          var population = new List<Individual>();  
6.          for (int i = 0; i < gp.populationParameters.populationSize && !timeoutInfo.ShouldTimeout; i++)  
7.          {  
8.              const bool forceFullyGrow = false;
9.              Node randomTree = gp.GenerateRandomTreeFromTypeOrReturnType<T>(gp.populationParameters.maxDepth, forceFullyGrow);  
10.             var ind = new Individual(randomTree);  
11.             await gp.EvaluateFitnessOfIndividual(ind);  
12.             population.Add(ind);  
13.         }            
14.         return population;  
15.     }  
16. }