This repository contains code to reproduce the results of our paper EvoLearner: Learning Description Logics with Evolutionary Algorithms.

In order to run all experiments, this repository contains SML-Bench.

The code of EvoLearner can be found in this folder.

• Ubuntu 18.04 LTS
• Python 3.6.9+ as python
• Java 8/11
• Apache Maven 3.6.0+
• 32GB RAM

Clone the repository:

git clone https://github.com/EvoLearnerOnto/EvoLearner.git

Then run:

./setup.sh

Alternatively, use the provided Dockerfile:

docker build -t evolearner .
docker run -it --rm --name=evolearner evolearner

When running the experiments below in the container, they will be written to the results folder in the container.

To make them available outside the container, you can mount a local directory:

docker run -it -v /path-to-local-directory:/sml-bench/results --rm --name=evolearner evolearner

To install Aleph follow the instructions here.

This is not required, so if Aleph is not installed the results of Aleph will just be missing.

To reproduce the results of EvoLearner, CELOE, OCEL, SPaCEL, Aleph (Table 3) run:

./reproduce_systems.sh

To reproduce the results of the ablation analysis of EvoLearner (Table 4) run:

./reproduce_ablation.sh

To reproduce the results of different variants of the random walk init (Table 5) run:

./reproduce_random_walk_variants.sh

To reproduce the results of the initialization methods (Table 7) run:

./reproduce_init_methods.sh

To reproduce the results of different mutation operators (Table 8) run:

./reproduce_mutation.sh

To reproduce the results of different settings for the maxT parameter (Table 9) run:

./reproduce_maxT.sh

To reproduce the results of different settings for the fitness function run:

./reproduce_fitness.sh

To reproduce the results of the F-measure over runtime experiment (Figure 3) run:

./reproduce_plot.sh

Afterward, the results can be found in the results folder.

Some solutions that were found by the systems for the Uncle learning problem:

EvoLearner

Male ⊓ ((∃ hasSibling.Parent) ⊔ (∃ married.(∃ hasSibling.Parent)))

• Perfect solution both on training and test data, short length

CELOE

(Son ⊓ (∃ hasSibling.Parent)) ⊔ ∃ married.Sister

• Short length, not 100% correct (Son vs. Male, Sister vs. hasSibling.Parent)

OCEL

Male ⊓ ((∃ hasSibling.Parent) ⊔ (∃ married.(Daughter ⊓ ∃ hasSibling.Parent)))

• Perfect solution on training and test data but a bit longer than necessary (one atomic concept too much: Daughter)

SPaCEL

(¬Female ⊓ (∃ hasSibling.Parent)) ⊔ (¬Female ⊓ ∃ married.(∃ hasSibling.Parent)))

• Perfect solution on training and test data but a bit longer than necessary (Male expressed as ¬Female, and ¬Female expressed two times)

The following table compares the final F1-measure of EvoLearner when running the complete algorithm with the F1-measure

directly after the initialization, so directly after the random-walk initialization without running the evolutionary algorithm afterward.

Showing the influence of different settings of the weight parameter of the fitness function.

(by how much the quality of an individual, i.e. concept, is weighted compared to its length)

• SML-Bench (used to run all experiments): https://github.com/SmartDataAnalytics/SML-Bench
• DL-Learner (CELOE and OCEL): https://github.com/SmartDataAnalytics/DL-Learner
• SParCEL: https://github.com/tcanvn/SParCEL
• DEAP: https://github.com/DEAP/deap
• Aleph: https://www.cs.ox.ac.uk/activities/programinduction/Aleph/aleph.html
• Owlready2: https://github.com/pwin/owlready2