SharpNEAT Roadmap

(Updated 22/12/2007)

Here are the current plans for SharpNEAT. The goals are mostly in order of highest to lowest priority from the top down.

Goal

Notes

Framework restructuring/rebuilding on .Net 2.0 This includes extensive use of Generics to simplify the framework, improve performance and provide a foundation for supporting CPPNs and other extensions.

Composition Pattern Producing Networks (CPPNs) CPPN is the indirect encoding devised by the EPLEX group at UCF.

David D'Ambrosio has done some work in this area and has released his CPPN enabled version of SharpNEAT (HyperSharpNeat) on the EPLEX web site.

New ANN model that defines neurons with multiple incoming connection points. This allows for multi-parameter activation functions. Traditional single activation function approach becomes a specific implementation of a more general ANN model. New activation functions can be more like GP functions, e.g. multiplication or performing routing (decision making).

Unify CPPNs, GP and ANN models into one generic model based only on graphs of procesing nodes. Genetic Programming (GP), and ANNs are really just restrictions on a model that has no restrictions to network topology or node activation function.

New front-end GUI The old GUI is tightly coupled to NEAT. The new one needs to be more flexible and extendable.

Revisit Function Regression. Specifically the WaveGenerator experiment. Quantify potential enhancement of NEAT gained from new ANN model(s) . Function regression is simple way of evaluating the power of a given neuro-evolution method.

Distributed (Sharp)NEAT Need to clarify which distributed model(s) to implement. Some form of island model seems likely as it should be easy to implement and should scale up nicely.

Also on this subject Wesley Tansey developed ParaSharpNeat (parallel SharpNEAT) while at Virginia Tech. ParaSharpNeat is available from Codeplex.

Tic-Tac-Toe co-evolution experiments The experiment code already exists. Retry it against new ANN model. General ongoing Investigations.

GECCO Competitions

Goal				Notes
Framework restructuring/rebuilding on .Net 2.0				This includes extensive use of Generics to simplify the framework, improve performance and provide a foundation for supporting CPPNs and other extensions.
Composition Pattern Producing Networks (CPPNs)				CPPN is the indirect encoding devised by the EPLEX group at UCF. David D'Ambrosio has done some work in this area and has released his CPPN enabled version of SharpNEAT (HyperSharpNeat) on the EPLEX web site.
New ANN model that defines neurons with multiple incoming connection points. This allows for multi-parameter activation functions.				Traditional single activation function approach becomes a specific implementation of a more general ANN model. New activation functions can be more like GP functions, e.g. multiplication or performing routing (decision making).
Unify CPPNs, GP and ANN models into one generic model based only on graphs of procesing nodes.				Genetic Programming (GP), and ANNs are really just restrictions on a model that has no restrictions to network topology or node activation function.
New front-end GUI				The old GUI is tightly coupled to NEAT. The new one needs to be more flexible and extendable.
Revisit Function Regression. Specifically the WaveGenerator experiment.				Quantify potential enhancement of NEAT gained from new ANN model(s) . Function regression is simple way of evaluating the power of a given neuro-evolution method.
Distributed (Sharp)NEAT				Need to clarify which distributed model(s) to implement. Some form of island model seems likely as it should be easy to implement and should scale up nicely. Also on this subject Wesley Tansey developed ParaSharpNeat (parallel SharpNEAT) while at Virginia Tech. ParaSharpNeat is available from Codeplex.
Tic-Tac-Toe co-evolution experiments				The experiment code already exists. Retry it against new ANN model. General ongoing Investigations.
GECCO Competitions