Using Artificial Life to Solve Problems

In this thesis, a method incorporating artificial life and genetic algorithms into a hybrid system for problem solving is presented. An ecosystem is created containing agents that live off rewriting solutions to problems. If an agent rewrites a solution so that it becomes a better solution, the agent gains some energy. This amount of energy is proportional to the improvement of the solution. The methodology for solving the problem is encoded into the system as the way the different species of agents organise themselves into a food web. The system learns to become more efficient by the presence of an evolutionary pressure. This pressure stems from the old agents that are regularly dying, and being replaced by new agents being born. Agents that are good at extracting energy from solutions, and hence are more helpful in finding the best solutions, have a higher probability of reproducing than other agents have. This ensures that the agents of the system will become better at solving the problems given to it. Several types of problems that are particularly difficult have been identified, and some improvements of the method are presented to circumvent these traps. The master's project has been unable to confirm that the method works. This is due to an error made during the design phase of the implementation. The dependencies between the agents were too weak, allowing agents that were dependent on each other to out-compete each other. The conclusion of the thesis is that it has not been possible to determine whether the method is viable or not. However, it is deemed that there is a potential for future research in the method and related areas. Foreword This master thesis and project is the final work needed for me to graduate from the Master of Computer Science program of the Royal Institute of Technology (KTH). Ever since I first heard about genetic algorithms, I have been fascinated by biologically motivated methods of programming. The method presented in this thesis is based on an idea I have had for a long time. The idea has gradually been improved over time and finally become the method presented in this thesis. I would like to thank my examiner Prof. Anders Lansner for allowing me to do this project and my supervisor Anders Sandberg for his enthusiasm and support. I would also like to thank Diane Rivo for helping me out with the language …