[x] I have opened an issue to discuss the proposed changes: Fix #314
[ ] I have modified/added tests to cover the new changes/features.
[ ] I have modified/added the documentation to cover the new changes/features.
[ ] I have ran the tests, checks, and code formater.
Proposed implementation
When computing the Khalil2016 features, some rows were considered active while they were not in the LP. Hence their LP index was -1, and the features of the last LP row were extracted for these rows. Luckily sometimes the last row was not itself active, hence the features were depending on the memory state (uninitialized memory), which resulted in non-deterministic features and was catched in #314
Pull request checklist
Proposed implementation
When computing the Khalil2016 features, some rows were considered active while they were not in the LP. Hence their LP index was -1, and the features of the last LP row were extracted for these rows. Luckily sometimes the last row was not itself active, hence the features were depending on the memory state (uninitialized memory), which resulted in non-deterministic features and was catched in #314