algebolic's Introduction

algebolic

Algebolic is a library for evolving mathematical expressions. You can use it to perform what's known as symbolic regression, where a symbolic mathematical expression is found that fits a given dataset. More generally, you can use it to evolve mathematical expressions that optimise any score function. The expressions are implemented so they can be very quickly evaluated, as can their derivatives which is useful in more advanced applications. We use it in our research lab, for example, to evolve laws of physics directly from data.

Algebolic is built on the darwin genetic algorithm framework.

Usage

Add it to your project.clj file:

[algebolic "1.0.0"]

Examples can be found in the ws/demo directory. The intro.clj worksheet is quite readable. With the others you might have to put in some thought/experimentation! You can view them on the web:

License

Distributed under the MIT licence.

algebolic's People

Contributors

Stargazers

Watchers

algebolic's Issues

Non-blocking random numbers

Current (rand) is slowing things down to effectively single-threaded speed, I think.

Improve SPEA2 performance

Tree manipulation and genetics tests

Some sort of hill-descent for the constants

Implement automatic differentiation in java interpreter

Expression simplification

Functions for examining multi-objective run results

Implement NSGA-II

Better tree generation algorithm

At the very least implement GROW. Probably also should do PTC2 or similar.

Improve tree-manipulation performance

Consistent score signs

Refactor metrics code

The metrics code doesn't seem right to me at the minute: there's some duplication, and things are spread around in an unpleasant way. I wonder whether something more like logging would be appropriate?

Think about individuals and state

At the moment scoring, fitness, selection etc all drop properties directly onto individuals. I'm not sure this is the best way to do it, as it could be confusing and lead to subtle bugs where something isn't cleared that should be. Should think about the best way to organise state information in the EA loop.

Implement 90-10 crossover

No idea whether it makes any difference, but it's what other people often do, so would be nice for replicating their results.

Size limiting in mutation and crossover

Ability to seed random number generator

Will probably need to switch all of the random calls. Not sure how this will work with multi-threading.

Differentiation of expressions

Implement SPEA2

Unary minus?

Having to write CA rules for both binary plus and binary minus is getting a bit tiresome. I wonder whether a unary minus would be a better idea. Or, indeed, whether you need a minus at all if you have negative constants.