TAG: Learning Timed Automata from Logs
Authors: Lénaïg Cornanguer, Christine Largouët, Laurence Rozé, Alexandre Termier3949-3958
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments compare our approach to the related work and demonstrate its merits. The first extensive experimental study and comparison with state-of-the-art algorithms for learning TA. We also provide an experimentation on real-world data from logs of TV programs. Experiment on synthetic data. Experiment on real-world data. |
| Researcher Affiliation | Academia | 1Inria 2Univ Rennes 3CNRS 4IRISA 5Institut Agro 6INSA Rennes |
| Pseudocode | Yes | Algorithm 1: TAG : P(TS(Σ, T)) A; Algorithm 2: merge : A A; Algorithm 3: split : A P(TS(Σ, T)) A |
| Open Source Code | Yes | an implementation of the algorithm is available1. 1TAG source code can be found here: https://gitlab.inria.fr/lcornang/tag/. |
| Open Datasets | Yes | We also carried out a real-data experiment to demonstrate TAG s ability to deal with real-world data and to produce interpretable models, using the logs of the programs of the Canadian TV channel CBC Windsor (Canadian Radio-television and Telecommunications Commission 2015). We first used the data of the Friday mornings of August 2020 (from 6:00 AM to 12:00 AM, one word per day). Then, we used the data of every day of July and August 2020 (Canadian summer school vacations months). The entries of the logs were summarized by their class (commercial message, promotion for a program ...) or category (program for children, news...) in case of a program. A word consists of the sequence of entries class/category and their duration for a day. |
| Dataset Splits | No | The paper describes how input samples (logs) are used for learning and how additional generated data is used for evaluation (testing), but it does not specify explicit train/validation/test splits of the input log data itself for the training process. |
| Hardware Specification | Yes | The runtime was measured on a Mac Book Pro with an Intel Core i9 processor clocked at 2,4 GHz and a memory of 16 Go 2667 MHz DDR4. |
| Software Dependencies | No | The paper states that 'TAG is implemented in Python, Timed k-Tail s author implementation (where k is hard coded to 2) is in Java and RTI+ in C++.' It mentions programming languages but does not provide specific version numbers for these languages or any libraries used. |
| Experiment Setup | Yes | RTI+ was executed with a significance value of 0.05 for the likelihood ratio test (default value). Tk T was executed with no nested events considerations (absent in our data) and no enlargement of the guards (default value). k is the only parameter of TAG. By controlling the length of the event sequences to compare, it allows tuning the trade-off between generalization and over-fitting of the model. If the input sample is exhaustive or if detecting wrong behavior is more important than having a small and easily interpretable model, k should be increased. Its default value is set to 2. |