NIMS Lab Evolutionary Computation Code Distributions

• Tangled program graphs (or TPG): provides a framework for the emergent discovery of complex organizations of programs through (symbiotic) co-operative coevolution. The resulting solutions are discovered under extremely modest computational requirements (single core/thread) and are typically orders of magnitude faster than commodity Deep learning solutions.
  • Stephen Kelly originally demonstrated TPG under the Arcade Learning Environment where this is available as a C++ code distribution.
  • Robert Smith developed a TPG distribution in Java for the ViZDoom first person shooter environment.
  • GEGELATI distribution represents a C++ TPG distribution and a nifty video tutorial of the TPG approach to reinforcement learning.
  • Ryan Amaral has a Python TPG distribution compatable with the OpenAI project and therefore the Atari suite of arcade titles.
• Hierarchical code reuse and task transfer: this is half of Stephen Kelly’s PhD thesis in which he demonstrated the ability to evolve policies for playing different tasks (e.g. possession and shooting on goal) and then learn how to redeploy these as a strategy for playing Half field offence. Issues of importance include the maintenance of diversity without recourse to task specific metrics. The framework is also deployed under Ms PacMan and the resulting policy demonstrated to reach the same quality as state-of-the-art results from Monte Carlo Tree Search and neuro-evolution. Code distribution comes with a couple of animations for illustrating the resulting policies in action.
• Symbiotic bid-based GP : Phython distribution originally developed by Jessica Pauli de Castro Bonson. Two versions available depending on whether you want classification or temporal sequence learning (most of the code is common). Both of these make use of dual diversity mechanisms (team diversity and normalized compression distance) in addition to the task specific fitness function (see Kelly and Heywood (2014)).
  • Classification tasks so good for multi-class classification from a single run.
  • Reinforcement learning tasks which includes a simple tic-tac-toe environment by way of a 'hello world' example task.
• Symbiotic bid-based GP : C++ distribution as originally developed by Peter Lichodzijewski. Being the 'original' SBB code base this assumes implicit fitness sharing as the diversity mechanism (no radii parameter necessary)
• Symbiotic bid-based GP : Java distribution (SSBJ) developed by Robert Smith. Also assumes implicit fitness sharing as the diversity mechanism.
  • This is a JAR file designed for participation in the FCUBE cloud computing challenge organized as part of the Evolutionary Computing for Big Data and Big Learning Workshop at ACM Genetic and Evolutionary Computing Conference (GECCO).
  • ... and a slightly updated version SBBJ-v02.1.rar
• Symbiotic Evolutionary Subspace Clustering : originally developed by Ali Vahdat.