An AIbase Approach to Destination Control in Elevators

1
An AI-base Approach to Destination Control in
Elevators

• Jana Koehler Daniel Ottiger
• Schindler Lifts Ltd.
• Presented by Jon Beckham

2
Challenges for the Industry

• The continuous pressure to lower construction
  costs of buildings requires that the core space
  occupied by an elevator installation be reduced
  and that transportation performance be
  significantly improved. (Lower shaft numbers
  while maintaining performance.)
• Increasing competition requires a diversification
  strategy to provide new and individually tailored
  services to passengers.

3
Normal Operation

• Two to eight cars
• Usually up and down call buttons
• Passengers usually dont know location of cars
• If not picked up immediately, usually press call
  button again. And again.
• Once on, passengers sometimes change their minds
  and request a different floor
• Passengers interact with each other, but are not
  aware
• Holding elevator doors is courtesy to one
  passenger, but not to the rest of the people
  trying to use the elevator.
• Ideal for passengers is short waiting time and
  short journey time

4
Criteria for Elevator Control Evaluation

• HC5
• Specifies the handling capacity of a group of
  cars within five minutes in terms of percentage
  of building population served.
• Times
• Average and maximum journey and waiting times.
• Humans care more about waiting time than journey
  time. Getting on an elevator in under a minute
  is more important than spending three minutes on
  the elevator.

5
Open Questions

• What is the objective function for a group
  dispatching algorithm? Usually a vague
  combination of waiting and journey times.
• How can a control system acquire additional
  information about passenger needs?
• How can the performance of a controller be
  improved?
• How can passenger interfaces be improved beyond
  the simple buttons?

6
Acquisition of Traffic Information

• Manual surveys
• Manually analyzed videotapes
• Counting system connected to call buttons
• Weight-sensing devices
• Computer vision technology

7
Traffic Patterns

• Up-peak
• Enter at lobby floor and request upwards
  transportation
• Down-peak
• Request downwards transportation to lobby from
  any floor
• Interfloor
• Request transportation from any floor to any
  other floor

8
The 80s

• Expert Systems!
• Given predefined patterns, passenger counts,
  rules from human lift experts, the system would
  transition into the hand-crafted optimal
  dispatching mode for the situation.
• Didnt work because
• Identifying the patterns didnt work.
• Human lift experts are busy people, they dont
  have time to update the rules all the time.

9
The Early 90s

• How about Fuzzy Logic?
• The traffic intensity was defined fuzzily, then
  fuzzy rules determined the predominant traffic
  pattern which in turn influenced elevator
  control.
• Didnt work because predicting traffic flow
  proved impossible.

10
The mid-90s

• Neural networks!
• Train on simulations, identify one out of five
  predefined traffic patterns.
• Same as expert systems in the 80s, didnt
  overcome problems of pattern-triggered rules.

11
A Little Bit Later

• OTISs fancier Neural Network
• Dispatching decision based on estimated remaining
  response time (RRT)
• Trained in simulation, attempts to minimize RRT
  errors
• Improved predictions up to 20 on average
• Simple perceptron isnt adequate representation
  of RRT

12
Yet Another Approach

• Neural Nets in a Reinforcement Learning framework
• Objective function minimize squared sum of
  waiting times
• 60,000 hours of training only on down-peak
  patterns
• Outputs of STOP or CONTINUE-DOWN
• Compared to comically simple policies, and not
  surprisingly proved to be better
• Building specific training makes this method
  pretty useless

13
Last One!

• Genetic Algorithms
• Fitness function weighted sum of waiting time,
  journey time, estimated passenger load per car
• The usefulness of genetic algorithms is not yet
  clear. Finding the right combination of specific
  crossover, mutation and selection methods
  yielding good dispatching decisions poses a
  challenge in this domain.

14
Ha, that wasnt the last one.

• Combinatorial Optimization of Travel Routes
• Minimin (single player minimax) lookahead search
  with alpha pruning
• One floor decision executed, re-search, repeat

15
Destination Control Systems

• Miconic-10 was introduced by Schindler in 1996.
  1500 elevators have been equipped with
  Miconic-10. Doubles HC5 of conventional
  dispatching algorithms. Yay!
• Ten digit keypad instead of traditional up/down
  call buttons.
• Allows for better prediction of travel routes,
  thus better utilization of lifts.

16
The Future of Elevators

• Access restrictions
• Oskar lives in the penthouse, doesnt want the
  factory workers to be able to go to his floor.
• VIP service
• Firefighters/medical personnel need to be able to
  immediately get control of elevators.
• Separation of passenger groups
• In hotels, dont want food service and
  housekeepers on the same lift for hygienic
  reasons.

17
Finally, the purpose of the paper

• The algorithmic methods used in the elevator
  industry have thus far not allowed for those
  functionalities to be integrated directly into
  the normal operation of a group of elevators.

18
Destination Control

• NP-hard
• The allocation problem with destination calls can
  be defined as follows
• Given a number n of destination calls with
  boarding floor b and exit floor e we wish to
  compute a totally ordered sequence of stops S
  such that each s corresponds to a given boarding
  or exit floor and where each b precedes the
  respective e.

19
Model Destination is

• Planning Problem?
• Scheduling Problem?
• Constraint Satisfaction Problem?

20
This is AI Planning

• So its most naturally a planning problem.
• Initial state is described by the current
  distribution of passengers and elevators.
• Goal state is any state satisfying that all
  passengers have been delivered to their
  destination floors.
• Available actions STOP, UP, DOWN, OPEN, CLOSE

21
Subproblems

• Static, offline optimization for one elevator
• Dynamic, online allocation problem for several
  cars
• Algorithm should allow new services to be added
  to destination control.
• Should be able to handle multi-deck elevators.

22
Physical System
23
Offline Problem

• The size of the search space is determined by
  stops, not by passenger volume.
• Real-time requirements are demanding. Must be
  able to compute instant allocations of passengers
  to cars. Must be able to quickly recompute
  travel routes to handle traffic changes. Upper
  time bound of 100ms.

24
The Algorithm

• Depth-first branch-and-bound modified for
  forward-checking from constraint reasoning.
• Allows for faster pruning of states in violation
  of service requirements.
• Computes optimal stop sequences for lifts with
  arbitrary numbers of decks.
• Prunes about 2/3 of states.
• Scales up past the biggest lift system.

25
Effectiveness of Pruning
26
Online Problem

• Auctioning methods used.
• Each elevator can communicate via asynchronous
  messaging supporting publish/subscribe mechanisms
  and allowing p2p communication between lift
  components.
• New agents can dynamically register. (Ad hoc
  networking)

27
Auction Method
28
Contract Net Protocol

• Broker receives offer requests from terminals,
  adds requests to the world model.
• Car driver responsible for executing the plans.
• Observer continuously updates the world model
  of the planner.
• Failure recovery monitors plan execution,
  diagnoses problems, initiates recovery actions.
• Drive executes start and stop commands.
• Doors execute opening commands.
• Configuration manager provides information
  about building layout.

29
Job Manager as Holon
30
Empirical Results
31
Problems

• With higher traffic, plans are replaced often,
  thus long plans unlikely to be executed.
• Solution, restrict plan length.
• Why use comprehensive plans?
• Why bother with figuring out optimal plans for
  everyone in the building instead of just the next
  optimal step?
• Often taking everything into account provides the
  means for an in-depth analysis of traffic.

32
Available Improvements

• Communication between passengers and elevators
  could be improved significantly.
• Current elevator systems constitute a significant
  waste of space. (One shaft per car.)