Site: Japan Air Lines
c/o JAL Creative Systems
12-38 Shibaura 4-chome
Tokyo 108, Japan
Date Visited: March 27, 1992
Mr. Sanshiro Shirahashi
Mr. Akira Mori
Mr. Yasunari Imamura
Mr. Mori, who championed the introduction of the flight crew scheduling system at JAL, discussed the history and development of that system, its architecture, user interface, etc.
Japan Air Lines (JAL) has two ESs in routing operations, both for planning and scheduling. The more important of the two is COSMOS/AI, for flight crew scheduling (see discussion below). The system was built by NEC on NEC's EXCORE tool, contains 45,000 frames, 220,000 lines of LISP code and 80,000 lines of C code. The system was developed over a period of 2.5 years and cost approximately $4 million. It was turned over to JAL for regular use in 1990. The primary benefits have been in cost reduction and quality improvement, but no quantitative results were given, except an estimate of six to seven years to pay back development costs. No personnel were laid off; with COSMOS/AI, JAL is able to schedule more planes with the same staff.
JAL has about 100 aircraft in their fleet (767, 747, 747-400, and DC-10s), and employs about 2,200 people as flight crew (pilots and flight engineers, etc.). Total cost of the system was 5 billion yen; the system has not quite paid for itself yet. Constraints on crew include training, government rules, union rules, and other customs.
Before the ES, 25 personnel were involved in flight crew scheduling. It took about six months to become expert in scheduling, assuming prior background in the constraints (about another six months to gain that knowledge). This has been reduced to two months total by the expert system.
The database on all flight information resides on an IBM 3090 mainframe. Data are updated by downloading from the host machine once a day at 3:00 a.m. JAL wrote a 3090 program which gathers changes, e.g., vacation days, etc.
The scheduling system runs on several UNIX workstations networked to the IBM system. An AI processor (NEC LISP machine) is attached to the workstations that process 747 schedules (busiest by far).
The basic function of the system is to make a schedule for the following month. This is due out on the 25th of each month.
The system uses the EXCORE frame-based tool -- LISP-based.
JAL does not yet have the capability to do reactive re-scheduling during the month of an operating schedule, but is working on it.
Formerly the 747 schedule took 20 days to prepare. This normally included considerable overtime. By using the ES, that time is reduced to 15 non-overtime days. Also, scheduling personnel have been reduced by about five (during a period when the flight load went up considerably).
JAL also claims that schedules are now better with fewer mistakes (actually claiming no mistakes at all after a monthly schedule has been accepted).
A toy prototype convinced management to proceed with the system. Acquisition of the knowledge was the hardest part.
The system became operational in February 1990.
The system is basically a heuristic one. Most constrained blocks (patterns of flight, stay in destination, return flight, required rest) are assigned first. Then blocks are shuffled (starting with least constrained) when problems are found. The system takes two to three hours to produce the complete schedule for the most complicated runs.
Currently optimization is not important (any correct schedule is good enough).
JAL began development in September 1987. NEC did the first prototype, but system work is now in-house.
An operations research approach was also tried. This was rejected as "too mechanical." The users did not like it because "human" constraints were ignored.
The system cannot completely finish a schedule. A human scheduling expert still makes final adjustments. After the schedule is finished by the humans, it is checked by ES to see that no constraints are violated (i.e., no mistakes).
Biggest virtue of the system is its maintainability. It is easy to add new rules about new planes, crews, etc. Two to three people are employed continually updating the scheduling KB.
JAL is also working on an aircraft cargo routing system (traveling salesman type problem), and on a career path system for flight crew members (i.e., tracking career milestones, etc.).
Although the system has not yet paid for itself, without it JAL may have needed more human schedulers, since JAL flights, crews, and planes have increased 20 percent, 10 percent, and 5 percent, respectively.
JAL management is very happy with AI (in the opinion of the group that we interviewed) and very proud of COSMOS/AI.
A paper on this system was presented at the First World Congress on Expert Systems in December 1991 (Onodera and Mori 1991). An earlier paper was presented at the 28th AGIFORS Symposium (Mori 1988).