Symbiotic Bid-based (SBB) GP source code

Code was developed by Peter Lichodzijewski and later extended for indirect indexing by John Doucette. Distribution is permitted for research purposes alone.

SBB Version 3 (2010): Purpose -- reinforcement learning

Relative to SBB Version 1 the following general differences appear:

Point population no longer expressed under a Pareto formulation.
- Assumes real-valued rewards, thus much more informative than the binary outcomes that makes a Pareto formulation so effective under classification domains.
- See Lichodzijewski and Heywood (2009) for some examples of this.
Asexual reproduction
- SBB version 1 employed crossover as one of the variation operators.
- Tends to result in the common team members being reinforced between the two parents.
- Experience under reinforcement learning tends to reward diversity with exploitation appearing under hierarchical model building.

Resource summary:

Online repository of the SBB project
Code distribution were this includes interfaces for:
- Monolithic versus latteral task decomposition: demonstrated under supervised learning tasks. Ultimate classification accuracies lower than SBB version 1 as no support for Pareto archiving for points and classifiers. However, the point of this paper was to contrast monolithic versus (latteral) cooperative task decomposition using exactly the same code base.
  - See Lichodzijewski, P., Heywood, M.I., Symbiosis, Complexification and Simplicity under GP, Genetic and Evolutionary Computation Conference, July 2010.
- Examples of reinforcement domain problems:
  - Doucette J.A., Lichodzijewski P., Heywood M.I. (2012) Hierarchical task decomposition through symbiosis in reinforcement learning. ACM Genetic and Evolutionary Computation COnference (GECCO), Nominated for Best Paper in ALife Track. (Case of the Acrobot Handstand task)
  - Kelly S., Lichodzijewski P., Heywood M.I. (2012) On run time libraries and hierarchical symbiosis. IEEE Congress on Evolutionary Computation (CEC)
  - Lichodzijewski, P., Doucette J.A. Heywood, M.I. (2011) Acrobot Handstand and Truck reversal Technical Report, Faculty of Computer Science, Dalhousie University.
  - Lichodzijewski, P. (2011) A Symbiotic Bid-Based (SBB) framework for problem decomposition using Genetic Programming PhD Thesis
  - Lichodzijewski, P., Heywood, M.I. (2010) The Rubik Cube and GP Temporal Sequence Learning, Genetic Programming Theory and Practice: VIII.

Version 1 (2008): Purpose -- Classification

Code is available using a regular indexing model assuming training data as an ascii text file (good for up to 64 attributes) or using indirect indexing (good for hundreds of thousands of attributes).

Regular SBB distribution
- Streaming data
  - Non-stationary streaming benchmark dataset
  - Heywood M. I. (2015) Evolutionary model building under streaming data for classification tasks: opportunities and challenges Genetic Programming and Evolvable Machines 16(3): 283-326
  - Vahdat A., Morgan J., McIntyre A. R., Heywood M.I, Zincir-Heywood A.N. (2015) Tapped Delay Lines for GP Streaming Data Classification with Label Budgets EuroGP. LNCS 9025:126-138.
  - Vahdat A., Atwater A., McIntyre A. R., Heywood M.I (2014) On the application of GP to streaming data classification tasks with label budgets. ACM GECCO (Companion): 1287-1294
  - Atwater A., Heywood M.I. (2013) Benchmarking Pareto archiving heuristics in the presence of concept drift: diversity versus age. ACM GECCO: 885-892.
  - Atwater A., Heywood M.I., Zincir-Heywood A.N. (2012) GP under streaming data constraints: A case for Pareto archiving? ACM GECCO.
- Batch (offline) classification
  - Lichodzijewski P., Heywood M.I. (2008) Managing Team-based Problem Solving with Sybolic Bid-based Genetic Programming. Proceedings of the Genetic and Evolutionary Computation Conference, GECCO'08, ACM Press
SBB with indirect addressing. Distribution as below.
- J. Doucette, P. Lichodzijewski, M. Heywood (2009) Evolving Coevolutionary Classifiers under Large Attribute Spaces. Genetic Programming Theory and Practice VII. Springer-Verlag
  - Comparison between SBB and C4.5 and MaxEntropy Embedded model of classification under Character Recognition and Bag-of-Words datasets from UCL repository. SBB source code and data sets provided below.
- J. Doucette, A. McIntyre, P. Lichodzijewski, M. Heywood (2012) Symbioic coevolutionary genetic programming: A benchmarking study under large attribute spaces. Genetic Programming and Evolvable Machines. 13(1).
  - Comparison between SBB and F+SVM model from LIBSVM tool box on the UCL Character Recognition and Bag-of-Words data sets. Develops a measure of complexity to provide comparison between SBB and SVM style of algorithm....
  - F+SVM required some modification, hence this is also supplied below although this only effects the SVM model selection phase and result reporting.

Component	Archive (zip)	File size
Coevolutionary embedded model:	Implementation	2.6 MB
Data sets formated for Coevolutionary mode (see UCI repository for originals):	SBB formated data	27 MB
F-score SVM filter model Basically LibSVM with modifications to the Python script to process large data files and report results in terms of CW-detection metric:	Implementation	118 KB
Data files from UCI with training/ text splits (formated for libSVM `fselect.py'):	F+SVM formated data	26.3 MB