The purpose of this program is to create an AI that can partially automate certain parts of the playtesting phase in the game development industry. The program is to reduce the time and cost of the playtesting phase. The solution program developed is chosen to be based on a neural network and a genetic algorithm. Furthermore a snake game is developed in order to test the programs functionality. Two methods of playtesting are successfully automated. However, it is concluded that the program can only be considered a proof of concept. This is due to the programs lack of ability to handle complexity, as well as the amount of implementation required should the program be effectively used on another game. Furthermore, if the program was ever to be used on more complex games, then the neural network and the genetic algorithm would need to be significantly expanded and large amount of computing power would be necessary.
A video of it completing snake: https://www.youtube.com/watch?v=LkzVnO3fc9U&list=PLlS5RhZexhv1ki2NWbjrGetReNjHF0yM1&index=8 Trainning UI: https://www.youtube.com/watch?v=EvhLvM_PPw4
The object GeneticAlgorithm makes it possible to create a genetic algorithm which is able to solve a unique problem.
Requirement 1: A class implementing the IFitnessCalculator which returns a class implementing the IFitnessInfo
The problem and a solution model should be defined in a class implementing the IFitnessCalculator interface. IFitnessCalculator requires the user to define the a method which returns a class implementing the interface IFitnessInfo containing the property Fitness, that represents how well the problem was solved. In the class that implements the interface IFitnessInfo, it is possible for the user to define uniqe properties to which the class implementing IFitnessCalculator can assign. The class implementing IFitnessInfo will be assigned to the property FitnessInfo in the Agent class, which makes it possible to access interesting information about the fitness calculation of that particular agent.
Requrement 2: Defining own implementation of classes implementing the interfaces ISelection, ICrossover, IMutation. Alternatively the user can use classes from the library implementing the interfaces with some of the most common methods:
ISelection
- TopPerformersSelector
- TopPerformersPlusAllTimeBestSelector (not impl.)
ICrossover
- OnePointCombineCrossover
IMutation
- RandomResettingMutator
Requirement 3: Creating and using the GeneticAlgorithm class
To create a new instance of the class GeneticAlgorithm it is necessary to specify which settings the algorithm should use by creating and passing and instance of the class GeneticSettings.
When the GeneticAlgorithm is instantiated, it contains a population of agents with randomized genes. To use the algorithm, the user has to invoke one or more of the following methods on the GeneticAlgorithm:
- CalculateFitness (takes a class implementing the IFitnessCalculator interface)
- MakeSelection (takes a class implementing the ISelctor interface. An exeption will be thrown if the user is trying to invoke the selection method on a population which has not been invoked with CalculateFitness first.
- MakeCrossover (takes a class implementing the ICrossover interface)
- MakeMutation (takes a class implementing the IMutator interface)
The user may have interest in accessing the following members:
Class Agent:
- TotalFitnessCalculations (represents number of fitness calculations in the lifetime of the genetic algorithm)
Class GeneticAlgorithm:
- AllTimeBestAgent (the best agent at the moment)
- AllTimeTopTenAgents (a list of the top ten performers)
- AllTimeTopTenAverageFitness (the average fitness value of the top ten performers above)
- Christian Damsgaard
- Magnus Lund
- Martin Lønne
- Simon Holst
- Shpend Gjelaa
- Sten Kirk Larsen
- Søren Hjorth Boelskifte