The code is based of paper Learning Search Space Partition for Black-box Optimization using Monte Carlo Tree Search.
LA-MCTS has three major components:
At each iteration, main loop builds the Monte Carlo search tree, selects next node, and samples on selected nodes.
Classifiers define rules to split a node, they also predict if a sample belongs to the node. Currently there are several builtin classifiers:
In these classifiers, some cluster algorithm is used to label the samples, then SVM is used to classify the samples. Builtin cluster algorithms include:
- KMeans
- Threshold
- Linear regression
A regressor is used to fit samples, then a threshold (median or mean) is used to separate them.
Samplers draw samples in node space. Currently builtin samplers include:
- Random sampler
- Bayesian sampler
- TuRBO sampler
- CMAES sampler
- Nevergrad sampler
Users may provide their own classifier and/or sampler by implementing Classifier and Sampler interface.
An example can be found at example_opt.py.
Detailed docs can found at here.
LA-MCTS is under CC-BY-NC 4.0 license.