Proposal

This is my summary of the paper that compares the quality and performance of job-shop scheduling solvers by Google and IBM. The paper can be found here.

compare the quality of and time to (both single and quad core) solutions by the two famous solvers: Google OR-Tools and IBM CPO
they compare these 2 against classic instances in the literature as well as a large scale benchmark with known optimal solutions
CPO in general performs better on both these types of instances
ORT performs better in the quad core setup
CPO is much superior when it comes to large scale problems

Summary

there are different types of scheduling problems: job-shop, flow-shop, open-shop etc
job-shop is the most popular among them and it tries to minimize the makespan of the whole process.
makespan = interval between the start of the earliest job and end of the latest job
till now this has been posed as a constraint satisfaction problem
recently it is being solved using LNS (Large Neighborhood Search) methods
they chose CP-SAT from OR-Tools for solving job-shop problem and kept the default optimizer from CPO
they only look at rectangular job-shop instances. Meaning:
- every job has to go through all the machines
- every job has number of ops equal to the number of machines
- every machine will have assigned a number of ops equal to the number of jobs
Both CPO and ORT use LNS as their technique
CPO uses portfolio strategies along with ML to converge to the best neighborhoods but ORT uses the simple variable/constraint selection techniques