CAVIAR Experiments

This experiment was based on the CAVIAR video dataset from:

Fisher, R.; Santos-Victor, J.; and Crowley, J. 2004. CAVIAR: Context aware vision using image-based active recognition. http://homepages.inf.ed.ac.uk/rbf/CAVIARDATA1/.

We used the version of the dataset as used by the authors of the OLED system:

Katzouris, N.; Artikis, A.; and Paliouras, G. 2016. Online learning of event definitions. Theory and Practice of Logic Programming 16(5-6):817–833

The goal is to learn rules that can be used to predict when two people are meeting.

In our AAAI 2020 paper, we also ran the ILASP (Inductive Learning of Answer Set Programs) system on the same dataset. The ILASP and FastLAS input files, together with scripts for running the experiments are available here.

The original dataset included event calculus axioms in the background knowledge. The learned rules were then about the initiating and terminating conditions of the meeting event. This relies on non-observational predicate learning, which is not supported by the current FastLAS system. We therefore derived the initiatedAt and terminatedAt examples from the data and gave these to FastLAS directly. For fairness, we also did the same for our comparison with ILASP (for which this simplification is unnecessary as ILASP does support non-observational predicate learning, and has previously been applied to the original version of the dataset).

To run the learning, run:

./run_learning.rb *system_name* > results_*system_name*.json

Where *system_name* should be replaced with either FastLAS or ILASP. To run the ILASP experiments, you will need to install ILASP.

To run the testing, run:

./run_test.rb *system_name*

This script will output a summary of the results.