This is a simple Rust library which implements linear genetic programming.
The library currently only supports single register input and output. This means the possible use cases would include evolving programs which encode single variable functions given a set of training data (i.e., (x, y) pairs).
This library uses tournament selection as well as self-adaptation of certain mutation parameters.
The data flow is similar to the classic LGP:
- Initialize population of random programs (only effective code, no semantic introns)
- Run data through each program and assign each program a fitness value based on MSE (mean-squared error)
- Perform tournament selection on population based on their fitness values
- Perform cross-over or mutation on winners of tournament selection
- Replace previous population with crossover/mutation resulting programs
- Repeat for fixed number of generations or until best fitness falls below set threshold
Currently needs a lot of work to make it faster. Many generations are needed to converge on exact solution or nearly exact solution.
- Re-write library to support variable input and output registers (this would make the use cases far more vast - e.g., time series)
- Perform testing to see which methods are most efficient (e.g., to allow introns or not), adding structural intron removal, etc
- Update and complete documentation
- Update and improve experience for user to set the configuration values
- TOP PRIORITY: Add sick GIF showing evolution of single-variable polynomial regression graph. Each frame is the graph of the current best program from generation 1 to the last generation.