DEGREE PROJECT, IN COMPUTER SCIENCE , FIRST LEVEL STOCKHOLM, SWEDEN 2015
A Comparison of Traditional Elevator Control Strategies
SHAYAN EFFATI & DONIA ALIPOOR
KTH ROYAL INSTITUTE OF TECHNOLOGY
A Comparison of Traditional Elevator Control Strategies
Degree Project in Computer Science (DD143X) Royal Institute of Technology
School of Computer Science and Communication
Shayan E↵ati Donia Alipoor shayane@kth.se donia@kth.se 0767 665 665 0735 250 894
Supervisor: Vahid Mosavat Examiner: ¨ Orjan Ekeberg
8th May 2015
Abstract
The purpose of this paper is to investigate which of a specific set of elevator control strategies that is the most time efficient from the pas- senger’s point of view in a specific office building. The paper first goes through five di↵erent strategies and approaches, followed by results from a simulation of some of the strategies and their combinations. From the test results, it can be concluded that one strategy works best for all pos- sible scenarios during a regular working day, both regarding the average waiting time for a passenger as well as the average travel time. The most optimal strategy implemented can be further optimized for more precise results, but it will not change the outcome of the comparison.
Referat
Syftet med denna rapport ¨ar att unders¨oka vilken av en upps¨attning hisskontrollstrategier som ¨ar den mest tidse↵ektiva ur passagerarnas syn- punkt. Unders¨okningen kommer ske i ett specifikt scenario i en kontors- byggnad. Rapporten g˚ar f¨orst igenom fem olika strategier och tillv¨aga- g˚angss¨att, f¨oljt av resultat fr˚an en datorsimulering av n˚agra av de n¨amnda strategierna och kombinationer av dessa. Fr˚an testresultaten kan slutsat- sen dras att en av strategierna funkar b¨ast f¨or alla m¨ojliga scenarion f¨or en vanlig arbetsdag, b˚ade med tanke p˚a genomsnittlig v¨antetid f¨or en pas- sagerare samt p˚a genomsnittlig resetid. Den mest optimala strategin som var implementerad kan optimeras ytterligare f¨or mer exakta resultat, men detta kommer inte ¨andra utg˚angen av j¨amf¨orelsen.
Contents
1 Introduction 5
1.1 Problem statement . . . 5
1.2 Scope . . . 5
2 Background 6 2.1 Elevator Control Strategies . . . 6
2.1.1 Collective Control . . . 6
2.1.2 Destination Control . . . 7
2.1.3 Zoning Approach . . . 7
2.1.4 Search-Based Approach . . . 7
2.1.5 Rule-Based Approach . . . 8
2.2 Passenger Patterns . . . 8
3 Method 9 3.1 Motion Control and Other Dynamics . . . 9
3.2 Simulated Passenger Pattern . . . 10
3.3 Implementation of Elevator Control Strategies . . . 11
3.3.1 Collective Control . . . 12
3.3.2 Zone Based Approach . . . 12
4 Results 14
5 Discussion 16
6 Conclusion 18
References 19
1 Introduction
Elevator control strategies are used in most elevators all over the world to deter- mine how elevators should behave. These strategies have been developed over a long time, and mostly all of them with the purpose of controlling the flow of passengers in an intelligent way. In a regular elevator, for example in an office or an apartment building have common patterns for how passengers will behave.
These patterns mostly depends on the time of day. In an office building, the passenger flow will be up-peak from the lobby, during the day mostly between di↵erent floors, and in the evening there will be down-peak traffic from di↵erent floors to the lobby.
Because of the long history of control strategies, elevators can have di↵erent behaviors in order to optimize the flow. This report will first provide background to the purpose of di↵erent control strategies, and then compare and analyze them to answer our problem statement.
1.1 Problem statement
In a regular office building the passenger flow of elevators can have huge impact on a company’s efficiency of manpower. The main objective of this paper is to analyze and compare several, as well as combinations of elevator control strategies. The results will determine which strategy will work best for an office building with two elevators and a common passenger flow. In this case the focus will be on minimizing waiting and travelling time for passengers using the elevator system.
This report will focus on an office building consisting of 10 floors, using 2 elevators, which have the maximum capacity of 10 passengers. The number of workers in the building is set to be 960 and a certain passenger pattern in the building is assumed and further described in section 3.2.
A simulator will be implemented in Java, and in turn used to simulate results based on several variables, including stochastic variables such as number of travelling passengers and desired destinations, and constants such as number of floors and elevator speed.
1.2 Scope
Even though researches in this field nowadays have a focus on machine learning strategies and search-based approaches, these kind of solutions makes a greater impact in building with an unpredictable passenger flow. Due to the fact that the passenger flow in an office building is relatively predictable, the investigated strategies in this report have been chosen accordingly. The focus in this report will rely on traditionally used strategies and approaches.
2 Background
Depending on what elements a particular building have when it comes to number of floors, passenger flow, and type of elevator system, a suitable operation and motion control system must be provided.
Motion control is the classification for the equipment related to the perfor- mance characteristics of an elevator, such as its speed, the time it takes for the elevator door to open and close etc. The operation control does not often a↵ect these factors. Furthermore, a distinction is made between motion control and operation control, which classifies all the electrical decisions designed into an elevator system to control the sequence of movements an elevator or elevator group will make in response to calls for service. [1]
In this section di↵erent strategies for operation control systems are described, and also how these are related to di↵erent motion control systems. Since this report focuses on a specific type of building using multiple car elevators, the background will be angled accordingly.
2.1 Elevator Control Strategies
Elevator control problems are stochastic, since there is uncertainty in the pas- senger patterns [2], when it comes to the passengers time of arrival, floor of arrival and desired destination floor. The methods used for elevator control strategies are often heuristic [3]. When it comes to control strategies, Rein- forcement Learning algorithms are ways of giving approximate solutions, by collecting data during time segments it is possible to find out when and where the passenger traffic will have its peak. The elevator system can further learn to adapt to the passenger patterns [1]. Other control strategies deals with the uncertainties using Fuzzy Logic. In contrast to traditional strict logic, that deals with terms and values such as ”true-false” and ”either-or”, Fuzzy Logic can work with more rough values such as ”true”, ”not true”, ”very true”, ”not very true”,
”more or less true” and so on [1].
Many algorithms are based on minimization of the remaining response time (RRT) for each passenger [3]. RRT is described as the calculated time it would take for each passenger to be picked up by an elevator considering all the current assignments it has [4]. The following subsections describes a set of di↵erent control strategies using di↵erent techniques. Some of which are operation control systems that can be implemented as they are described, and some which are approaches when implementing a control strategy.
2.1.1 Collective Control
One of the earliest strategies for elevator control systems is the collective control strategy. Here each floor is equipped with buttons representing the directions up and down and the car keeps track of all the calls made in the same direction, collecting all passengers going in the same direction. The car then reverse and collects all passengers going in the opposite direction. [1]
The collective control strategy can also be put to use using multi-car ele- vators. When it comes to two cars operating in parallel one of the two cars is called the home-landing car while the other car is free. Which car is the one or the other may di↵er from time to time. The home-landing car is assigned to collect any passenger calling from the lobby or above it and below the free car.
If the home-landing car is commanded to go to a floor above the free car, the free car switches to become the home-landing car. In a situation where many calls are made after each other, the cars may start ”leapfrogging”, resulting in them going in the same direction. This is no better than having a larger single car using the collective control strategy and the use of two coordinated cars looses its purpose. This phenomenon is called ”bunching” and is a undesirable e↵ect for a multi-car elevator system. [1]
2.1.2 Destination Control
A strategy to optimize elevator control for passengers sharing the same destina- tion direction, but maybe not the same destination floor, is to have passengers entering their desired destination floor before entering an elevator car. This is called Destination Control and is a desired strategy for larger buildings where only the desired direction is not sufficient information for optimizing which el- evator to dispatch. [5]
2.1.3 Zoning Approach
In a zoning approach the goal is to reduce the number of stops each elevator makes, using several elevators which are assigned to serve a certain zone. For example some may serve lower floors, while others may serve higher floors.
This approach may be used in heavy traffic when there’s a spread in the chosen destinations, making sure that several elevators doesn’t answer the same calls at the expense of providing the passengers flexibility. Depending on the passenger flow and floors of importance, the zones should be divided accordingly. The number of zones often corresponds to the number of existing elevators. [1] This approach is comparable to many subway systems, which uses the principle of dividing the system into di↵erent lines, corresponding the zones in these kinds of elevator systems.
2.1.4 Search-Based Approach
The control strategies that uses a search-based approach, assigns a car with the choice that optimizes some criterion, such as the average waiting time or traveling time. In contrast to the former introduced elevator control strategies which all are based on greedy algorithms, this approach may search through the space of possible car assignments, which in this case makes it a non-greedy algorithm. A greedy algorithm decides for a car assignment as soon as it finds one that suits, while a non-greedy algorithm postpone their assignments or reconsider them if they might receive some other information about additional
calls or passenger destinations. Greedy algorithms are often less flexible than the non-greedy, but on the other hand they require less computation time and may execute faster. [6]
2.1.5 Rule-Based Approach
The Rule-Based approach refers to elevator control systems that uses condi- tional statements. In some sense all the control strategies described so far uses conditional statements; ”if situation, then action”. But the rule-based approach that is described by Cristes and Barto [6] uses fuzzy-logic and is often used in Artificial Intelligence. There lies a lot of difficulty in arranging and incorporat- ing di↵erent control strategies that would suit well for an elevator system under various passenger patterns. In order to shorter waiting and traveling times un- der these kind of conditions experts came up with using fuzzy control rules to solve the problem. The fuzziness i used in the if-part of the conditional state- ments. An example of a logical rule is described in Crites and Barto’s report as following:
if there is a hall call registered on an upper floor and there are a large number of cars ascending towards the upper floors then
assign one of the ascending cars on the basis of estimated time of arrival;
end
The rule searches for future car positions and probable future hall calls and assigns cars based on the estimated time of arrival immediately. This particular rule can be seen as a combination of a greedy search-based algorithm as was mentioned in Section 2.1.4, and a rule-based algorithm.
2.2 Passenger Patterns
Elevator control strategies are based on passenger patterns. It is of interest to analyse in which patterns passenger arrives and travels during the course of a day, in order to find the most suitable strategy. In this report focus will lie on analysing so called up-peak traffic and down-peak traffic. Up-peak traffic is characterized by having a single arrival floor and many destinations. For example, in an office building this would occur during the beginning of office hours. Many will arrive at the lobby of the building, with the purpose of going to di↵erent office floors. Down-peak traffic is characterized by having many arrival floors and one destination floor. This would be when office hours comes to an end, and people are heading home from their offices. The passenger patterns in up-peak, respective down-peak traffic is illustrated in Figure 1.
Figure 1: Demonstrating up-peak traffic in left picture and down-peak traffic in right picture.
3 Method
The best method for this project is to implement a simulator with the option of choosing what strategy to test. Choosing one simulator with di↵erent strategies, in contrast to one simulator for each strategy, is a good way to make sure that di↵erent implementations will not have di↵erent time measurements other than the time to run each algorithm.
Time optimization for runtime or compilation will not be an important fac- tor. When C may be the optimized programming language for time optimiza- tion, Java will be the programming language to be used for the simulator. The choice is based on the authors’ better experience in the language.
The simulator will give time measurements for each strategy, and all results can after that be compared with each other. This way the simulator doesn’t have to be optimized for being as fast as possible. The purpose is to find the most suitable strategy among the ones simulated.
3.1 Motion Control and Other Dynamics
The relevant motion control for this study is the speed of the elevators. The speed of the elevator is divided into floor time, stop time and turn time. Other
dynamics taken into concern are the boarding time and the car capacity. The variables listed in Table 1 are constant through all the simulations.
Variable Description Value
Floor time The time it takes to travel one floor at maximum speed.
1.5s Stop time The time it takes to decelerate,
stop, open and close doors and accelerate again.
7s
Boarding time The time it takes for one passen- ger to enter and exit the car.
1s Car capacity The number of passengers a car
can take.
10 passengers
Table 1: The constant variables of the elevator system
3.2 Simulated Passenger Pattern
In this investigation a specific passenger pattern is assumed. The elevators are assumed to be active during 10 hours, which corresponds to the working hours of the office. The first 2 hours is set to be the rush hours, meaning that we have heavy up-peak traffic during these hours. The last 2 hours are also rush hours, but in the opposite direction and have the down-peak characteristics. During the rush hours all 960 workers get to their offices, respective get out of their offices.
The passengers arrival is always constant. During rush hours a passenger arrives every 8 seconds, while during regular working hours a passenger arrives every 30 seconds. During up-peak 90% of the arrived passengers travels in a upward direction from the lobby, and 10% travels randomly in the rest of the building.
During down-peak 90% of the passengers travels in a downward direction to the lobby, and 10% randomly in the rest of the building. During the remaining 6 working hours people arrives and travels to random floors including the lobby.
The assumed passenger pattern suits this investigation due to the fact that an office building has a specific amount of passengers. The passengers also has specific destinations everyday when they arrive at the lobby every morning, namely their office floor. Likewise, each passenger has a specific arrival floor in the evening, assuming that all workers start and end their day in their offices.
Furthermore, a specific passenger pattern like this can be used to compare dif- ferent strategies in this type of building. Using stochastic variables would be an alternative, but such investigations need more collecting of data in order to identify a pattern.
3.3 Implementation of Elevator Control Strategies
A turn-based method is the chosen method for keeping track on time. Each turn represents 0.1s in real life. The values presented in 3.1 is translated to turns in the simulator and the simulator waits the set amount of turns before moving on. An implemented passenger class keeps track on both how many turns the passenger has been waiting for an elevator, as well as how many turns the passenger has been travelling in an elevator.
The elevator system simulated is in a building of 10 floors and a lobby, using 2 elevator cars. A class implementing an elevator system keeps track on:
• a set of elevators
• how many passengers that have been transferred to each floor
• passenger patterns and flow
• a collective passenger queue for all elevators
A class implementing an elevator has in turn control over:
• the passenger capacity
• passengers in the elevator
• the current state of movement
• how many turns each action takes
For each turn, the system checks if any of the elevators have any turns left to wait. If it does, the system adds to each passenger waiting time how many turns it needs to wait. If not, the system checks if an elevator is in transit, and then if it is in transit because of passengers on board or if it is in transit because the elevator is assigned to one or several waiting passenger(s). If in transit, it moves one floor, counting how many turns it now has to wait to take account of the time it takes to move. If the elevator has either reached a destination or if it has reached a floor with a waiting passenger with the same desired direction as the elevator is heading, it stops. While stopping, the elevator drops o↵ possible passengers that have reached their destination, as well as boarding possible passengers that have the same desired direction as the elevator is heading.
For each action, it waits out the amount of turns every action costs, therefore giving the passenger class an opportunity of counting how long a passenger has been waiting or travelling.
The algorithm to determine the action for each elevator for each turn works as following:
if no turns left to wait out and the elevator car is or should be in transit then
if a passenger is has a desired direction matching the elevators direction or the elevator has reached a floor matching a passengers destination then
stop the car and drop o↵ the passenger(s) and/or pick up the passenger(s)
end else
move one floor end
end else
wait one turn end
An elevator car should be in transit when there is a passenger on board or there is a passenger in the queue waiting for an elevator.
3.3.1 Collective Control
If the elevator is empty, it starts by picking up the first passenger in the queue.
If a passenger with the same desired direction presses the button while the elevator is headed toward that direction, that passenger is picked up as well.
The elevator keeps on going in this direction until all passengers wanting to go in the same direction are dropped o↵, after which the elevator turns direction if a such request exists. The elevator system decides which elevator should get which passenger based on a series of factors:
• the current directions of the elevators
• the elevator closest to the passenger is prioritized
• the elevator should not be filled 3.3.2 Zone Based Approach
The zone based approach divides the building into two zones. One of the two elevators covers the first zone, which comprises the lobby and the floors 1-5.
The other elevator covers the second zone, which comprises the lobby and the floors 6-10. Before entering the elevator, the system uses Destination Control to determine which elevator to send depending on which zone the passenger desires. While travelling inside the zones, the system uses collective control.
In our case there are two scenarios of how a passenger can travel:
1. A passenger calls on an elevator in one and has a desired destination inside the same zone. Here the passenger is assigned to the elevator designated for that zone.
2. A passenger calls on an elevator in one zone and has a desired destination inside the other zone. If the passenger is in the lobby it is assigned to the elevator designated for the other zone directly. If the passenger is on any other floor in one zone, it needs to travel to the lobby before switching to the elevator designated for the other zone.
4 Results
The simulation represents a working day, and the results gotten from the sim- ulation is presented in Table 2. The time is translated from the unit ”turns”
mentioned in the method.
Collective Control Zoning Approach
up random down up random down
Average waiting time Elevator 1 27.4s 15.0s 9.8s 35.5s 30.2s 18.9s Elevator 2 23.0s 14.4s 9.3s 43.8s 27.9s 24.7s Average travel time Elevator 1 14.4s 14.0s 16.4s 12.3s 19.3s 16.8s Elevator 2 13.3s 14.7s 17.8s 20.6s 20.0s 23.2s
Table 2: Results from simulation of Collective Control and Zone Based control strategies.
As seen in the table, using only collective control as a strategy for the elevator system gives better results in most possible scenarios. The zone based approach gives an insignificant advantage during the rush hours, since the elevators can focus on a smaller amount of possible destinations, and therefore allowing stops in the lobby more often, and there picking up waiting passengers more often.
Since the average waiting time is mostly doubled in all scenarios, the zone based approach is not preferred over using only collective control.
During the morning rush hour, the zone based approach increased the av- erage waiting time with 57%. During working hours, the zone based approach increased the average waiting time with almost 98%, since it is highly likely that passengers need to switch between zones during midday. During the afternoon rush hour, the zone based approach increased the average waiting time with 128%.
Total floors traveled
Collective Control Zoning Approach
up random down up random down
Elevator 1 2642 5379 1585 2812 4192 1396 Elevator 2 2715 4565 1690 2863 4383 1961
Table 3: Travel statistics under di↵erent passenger patterns from the simulation.
Collective Control Zoning Approach
Total trips 2640 2640
Total floors traveled 12850 13148
Floors/trip 4 4
Table 4: Average travel statistics from the simulation.
5 Discussion
What was learned from creating a simulator and comparing two di↵erent con- trol strategies, is that zoning in general has longer waiting and travelling time compared to basic collective control. Since the system only have two elevators in use, using zoning makes the two elevators seem like only one, especially in the up-peak and down-peak rush hours. If the demand is high on only one of the zones, it creates longer waiting times for the elevator responsible for that zone, since it’s the only elevator in use. It also makes the other elevator redundant when it can’t be used even though it’s heavy traffic. This is reflected in the results in the zoning approach for elevator 2, which is responsible for zone 2, compromised of the lobby and the floors 6-10. The long waiting time is probably due to the elevator being busy traveling all the floors in zone 2, before returning to the lobby. Even though elevator 1 exists, it is redundant for when a flow of passengers with desired destinations in zone 2 arrive close to each other.
According to Sorsa, Hakonen & Siikonen, considering the elevator handling capacity, up-peak traffic is the most demanding type of traffic.[7] This reflects our results, since even though the up-peak and down-peak rush hours are both rush hours, the up-peak period has significantly more waiting time. This is due to the ine↵ectiveness of an elevator that transfers passengers to di↵erent destination floors, with no passengers to return on the way down to the lobby.
This also reflects why the down-peak travel time is equal or a bit greater than the up-peak, since the elevator stops on the way down for passengers heading towards the same destination, adding more and more time for each passenger already in the elevator car.
There is a lot that can be done in order to optimize the results of what strategies are the most time efficient and make them more accurate. For exam- ple, using stochastic variables is one thing. Usually the kind of implementations using stochastic variables are used in control strategies based on reinforcement learning. That is, making a program that collects data of the travelling passen- gers and identify the actual pattern in the specific building. Such program would make the operation control smarter with time, as more data is collected and fur- ther get a clearer picture of which strategy is the most suitable for the specific passenger pattern. Since passenger flows in office buildings are relatively pre- dictable, approaches based on reinforcement learning would not generate more specific simulated passenger patterns.
In this report the assumed passenger pattern obviously will have an impact on the result. As can be seen in Table 2, the results di↵ers for same strategy in the three di↵erent pattern columns, thus the pattern has an impact on the data output.
In buildings with similar characteristics to the passenger pattern, the method used in this report can be reused. The variables when it comes to number of floors, number of workers, number of elevators, how high the pressure is during rush hours in up-peak respectively down-peak traffic, can easily be replaced with other values. Furthermore, the method can be used for other buildings with predictable passenger pattern, meaning not only office buildings, but also
buildings such as apartment buildings, where passengers in great extent only travel between the lobby and their floor. This report uses a specific flow, but any other flow can be studied and implemented using this report’s method, and further investigate which strategy that works best for that flow.
6 Conclusion
We have investigated how to solve the problem of which elevator control strategy to use for a single office building with 10 floors and 2 elevator cars. For the scenario described in the problem statement, and with the results produced by implementing the strategies and approaches covered in this paper, we can conclude that collective control is the most optimal strategy. A zoning approach only leaves the elevators ine↵ective for rush hours. The simulator developed for this paper can be optimized further, with a possibility to implement a search based approach and to implement a stochastic passenger flow. We believe that such optimizations only gives more thorough simulation results, but gives no change in the outcome of the comparison of the di↵erent strategies.
References
[1] Bob Caporale, Robert S Caporale, and George R Strakosch. The vertical transportation handbook. Hoboken John Wiley & Sons, Inc., 2010.
[2] Gina Carol Barney. Elevator Traffic Handbook. Routledge, 2003.
[3] Daniel Nikovski and Matthew Brand. Marginalizing out future passengers in group elevator control. In Proceedings of the Nineteenth Conference on Uncertainty in Artificial Intelligence, UAI’03, pages 443–450, San Francisco, CA, USA, 2003. Morgan Kaufmann Publishers Inc.
[4] B.A. Powell and J.N. Williams. Elevator dispatching based on remaining response time, September 8 1992. US Patent 5,146,053.
[5] Shunji Tanaka, Yukihiro Uraguchi, and Mituhiko Araki. Dynamic optimiza- tion of the operation of single-car elevator systems with destination hall call registration: Part i. formulation and simulations. European Journal of Operational Research, 167(2):550 – 573, 2005.
[6] Robert H Crites and Andrew G Barto. Elevator group control using multiple reinforcement learning agents. Machine Learning, 33:235–262, 1998.
[7] Janne Sorsa, Henri Hakonen, and Marja-Liisa Siikonen. Elevator selection with destination control system. In A. Lustig, editor, Elevator Technology 15, Proceedings of Elevcon 2005, number 15 in Elevator Technology, pages 202–211. IAEE, 2005.
