service engineering: data-based science & teaching in support … · 2019. 11. 26. · 5....
TRANSCRIPT
Service Engineering:Data-Based Science & Teaching
in support of Service Management
Avishai Mandelbaum
Technion, Haifa, Israel
http://ie.technion.ac.il/serveng
Based on joint work with Sergey Zeltyn, . . .
Technion SEE Center / Lab: Paul Feigin, Valery Trofimov, RA’s, . . .
1
Main Messages
1. Simple Useful Models at the Service of Complex Realities.Note: Useful must be Simple; Simple often rooted in Deep analysis.
2. Data-Based Research & Teaching is a Must & Fun.Supported by DataMOCCA = Data MOdels for Call Center Analysis.Initiated with Wharton, developed at Technion, available for adoption.
3. Back to the Basic-Research Paradigm (Physics, Biology, . . .):Measure, Model, Experiment, Validate, Refine, etc.
4. Ancestors & Practitioners often knew/apply the “right answer":simply did/do not have our tools/desire/need to prove it so.Supported by Erlang (1915), Palm (1945),..., thoughtful managers.
5. Scientifically-based design principles and tools (software),that support the balance of service quality, process efficiency andbusiness profitability, from the (often-conflicting) views ofcustomers, servers, managers: Service Engineering .
2
Background Material (Downloadable)
I Technion’s ‘‘Service-Engineering" Course (≥ 1995):http://ie.technion.ac.il/serveng
I Gans (U.S.A.), Koole (Europe), and M. (Israel):“Telephone Call Centers: Tutorial, Review and ResearchProspects." MSOM, 2003.
I Brown, Gans, M., Sakov, Shen, Zeltyn, Zhao:“Statistical Analysis of a Telephone Call Center: AQueueing-Science Perspective." JASA, 2005.
I Trofimov, Feigin, M., Ishay, Nadjharov:"DataMOCCA: Models for Call/Contact Center Analysis."Technion Report, 2004-2006.
I M. “Call Centers: Research Bibliography with Abstracts."Version 7, December 2006.
3
The First Prerequisite: Data & Measurements
Empirical “Axiom": The data one needs is never there for one touse – always problems with historical data (eg. lacking,contaminated, averaged, . . .)
Averages Prevalent.But I need data at the level of the Individual Transaction: For eachservice transaction, its operational history – time-stamps of events.(Towards integrating with marketing / financial history.)
Sources: “Service-floor" (vs. Industry-level, Surveys, . . .)
I Administrative (Court, via “paper analysis")I Face-to-Face (Bank, via bar-code readers)I Telephone (Call Centers, via ACD)I Future: Hospitals (via RFID)
4
Measurements: Face-to-Face Services23 Bar-Code Readers at a Bank Branch
Bank – 2nd Floor Measurements
5
Measurements: Telephone Call-by-Call Data (Log-File)
Telephone Service: Call-by-Call Data
vru+line call_id customer_id priority type date vru_entry vru_exit vru_time q_start q_exit q_time outcome ser_start ser_exit ser_time server
AA0101 44749 27644400 2 PS 990901 11:45:33 11:45:39 6 11:45:39 11:46:58 79 AGENT 11:46:57 11:51:00 243 DORIT
AA0101 44750 12887816 1 PS 990905 14:49:00 14:49:06 6 14:49:06 14:53:00 234 AGENT 14:52:59 14:54:29 90 ROTH AA0101 44967 58660291 2 PS 990905 14:58:42 14:58:48 6 14:58:48 15:02:31 223 AGENT 15:02:31 15:04:10 99 ROTH
AA0101 44968 0 0 NW 990905 15:10:17 15:10:26 9 15:10:26 15:13:19 173 HANG 00:00:00 00:00:00 0 NO_SERVER
AA0101 44969 63193346 2 PS 990905 15:22:07 15:22:13 6 15:22:13 15:23:21 68 AGENT 15:23:20 15:25:25 125 STEREN
AA0101 44970 0 0 NW 990905 15:31:33 15:31:47 14 00:00:00 00:00:00 0 AGENT 15:31:45 15:34:16 151 STEREN
AA0101 44971 41630443 2 PS 990905 15:37:29 15:37:34 5 15:37:34 15:38:20 46 AGENT 15:38:18 15:40:56 158 TOVA
AA0101 44972 64185333 2 PS 990905 15:44:32 15:44:37 5 15:44:37 15:47:57 200 AGENT 15:47:56 15:49:02 66 TOVA
AA0101 44973 3.06E+08 1 PS 990905 15:53:05 15:53:11 6 15:53:11 15:56:39 208 AGENT 15:56:38 15:56:47 9 MORIAH
AA0101 44974 74780917 2 NE 990905 15:59:34 15:59:40 6 15:59:40 16:02:33 173 AGENT 16:02:33 16:26:04 1411 ELI
AA0101 44975 55920755 2 PS 990905 16:07:46 16:07:51 5 16:07:51 16:08:01 10 HANG 00:00:00 00:00:00 0 NO_SERVER
AA0101 44976 0 0 NW 990905 16:11:38 16:11:48 10 16:11:48 16:11:50 2 HANG 00:00:00 00:00:00 0 NO_SERVER
AA0101 44977 33689787 2 PS 990905 16:14:27 16:14:33 6 16:14:33 16:14:54 21 HANG 00:00:00 00:00:00 0 NO_SERVER
AA0101 44978 23817067 2 PS 990905 16:19:11 16:19:17 6 16:19:17 16:19:39 22 AGENT 16:19:38 16:21:57 139 TOVA
AA0101 44764 0 0 PS 990901 15:03:26 15:03:36 10 00:00:00 00:00:00 0 AGENT 15:03:35 15:06:36 181 ZOHARI
AA0101 44765 25219700 2 PS 990901 15:14:46 15:14:51 5 15:14:51 15:15:10 19 AGENT 15:15:09 15:17:00 111 SHARON
AA0101 44766 0 0 PS 990901 15:25:48 15:26:00 12 00:00:00 00:00:00 0 AGENT 15:25:59 15:28:15 136 ANAT AA0101 44767 58859752 2 PS 990901 15:34:57 15:35:03 6 15:35:03 15:35:14 11 AGENT 15:35:13 15:35:15 2 MORIAH
AA0101 44768 0 0 PS 990901 15:46:30 15:46:39 9 00:00:00 00:00:00 0 AGENT 15:46:38 15:51:51 313 ANAT
AA0101 44769 78191137 2 PS 990901 15:56:03 15:56:09 6 15:56:09 15:56:28 19 AGENT 15:56:28 15:59:02 154 MORIAH
AA0101 44770 0 0 PS 990901 16:14:31 16:14:46 15 00:00:00 00:00:00 0 AGENT 16:14:44 16:16:02 78 BENSION
AA0101 44771 0 0 PS 990901 16:38:59 16:39:12 13 00:00:00 00:00:00 0 AGENT 16:39:11 16:43:35 264 VICKY
AA0101 44772 0 0 PS 990901 16:51:40 16:51:50 10 00:00:00 00:00:00 0 AGENT 16:51:49 16:53:52 123 ANAT
AA0101 44773 0 0 PS 990901 17:02:19 17:02:28 9 00:00:00 00:00:00 0 AGENT 17:02:28 17:07:42 314 VICKY
AA0101 44774 32387482 1 PS 990901 17:18:18 17:18:24 6 17:18:24 17:19:01 37 AGENT 17:19:00 17:19:35 35 VICKY
AA0101 44775 0 0 PS 990901 17:38:53 17:39:05 12 00:00:00 00:00:00 0 AGENT 17:39:04 17:40:43 99 TOVA
AA0101 44776 0 0 PS 990901 17:52:59 17:53:09 10 00:00:00 00:00:00 0 AGENT 17:53:08 17:53:09 1 NO_SERVER
AA0101 44777 37635950 2 PS 990901 18:15:47 18:15:52 5 18:15:52 18:16:57 65 AGENT 18:16:56 18:18:48 112 ANAT
AA0101 44778 0 0 NE 990901 18:30:43 18:30:52 9 00:00:00 00:00:00 0 AGENT 18:30:51 18:30:54 3 MORIAH
AA0101 44779 0 0 PS 990901 18:51:47 18:52:02 15 00:00:00 00:00:00 0 AGENT 18:52:02 18:55:30 208 TOVA
AA0101 44780 0 0 PS 990901 19:19:04 19:19:17 13 00:00:00 00:00:00 0 AGENT 19:19:15 19:20:20 65 MEIR
AA0101 44781 0 0 PS 990901 19:39:19 19:39:30 11 00:00:00 00:00:00 0 AGENT 19:39:29 19:41:42 133 BENSION AA0101 44782 0 0 NW 990901 20:08:13 20:08:25 12 00:00:00 00:00:00 0 AGENT 20:08:28 20:08:41 13 NO_SERVER
AA0101 44783 0 0 PS 990901 20:23:51 20:24:05 14 00:00:00 00:00:00 0 AGENT 20:24:04 20:24:33 29 BENSION
AA0101 44784 0 0 NW 990901 20:36:54 20:37:14 20 00:00:00 00:00:00 0 AGENT 20:37:13 20:38:07 54 BENSION
AA0101 44785 0 0 PS 990901 20:50:07 20:50:16 9 00:00:00 00:00:00 0 AGENT 20:50:15 20:51:32 77 BENSION
AA0101 44786 0 0 PS 990901 21:04:41 21:04:51 10 00:00:00 00:00:00 0 AGENT 21:04:50 21:05:59 69 TOVA
AA0101 44787 0 0 PS 990901 21:25:00 21:25:13 13 00:00:00 00:00:00 0 AGENT 21:25:13 21:28:03 170 AVI
AA0101 44788 0 0 PS 990901 21:50:40 21:50:54 14 00:00:00 00:00:00 0 AGENT 21:50:54 21:51:55 61 AVI
AA0101 44789 9103060 2 NE 990901 22:05:40 22:05:46 6 22:05:46 22:09:52 246 AGENT 22:09:51 22:13:41 230 AVI
AA0101 44790 14558621 2 PS 990901 22:24:11 22:24:17 6 22:24:17 22:26:16 119 AGENT 22:26:15 22:27:28 73 VICKY
AA0101 44791 0 0 PS 990901 22:46:27 22:46:37 10 00:00:00 00:00:00 0 AGENT 22:46:36 22:47:03 27 AVI
AA0101 44792 67158097 2 PS 990901 23:05:07 23:05:13 6 23:05:13 23:05:30 17 AGENT 23:05:29 23:06:49 80 VICKY
AA0101 44793 15317126 2 PS 990901 23:28:52 23:28:58 6 23:28:58 23:30:08 70 AGENT 23:30:07 23:35:03 296 DARMON
AA0101 44794 0 0 PS 990902 00:10:47 00:12:05 78 00:00:00 00:00:00 0 HANG 00:00:00 00:00:00 0 NO_SERVER
AA0101 44795 0 0 PS 990902 07:16:52 07:17:01 9 00:00:00 00:00:00 0 AGENT 07:17:01 07:17:44 43 ANAT
AA0101 44796 0 0 PS 990902 07:50:05 07:50:16 11 00:00:00 00:00:00 0 AGENT 07:50:16 07:53:03 167 STEREN
6
Averages Prevalent
ACD Report: Health Insurance
Time Calls Answered Abandoned% ASA AHT Occ% # of agentsTotal 20,577 19,860 3.5% 30 307 95.1%8:00 332 308 7.2% 27 302 87.1% 59.38:30 653 615 5.8% 58 293 96.1% 104.19:00 866 796 8.1% 63 308 97.1% 140.49:30 1,152 1,138 1.2% 28 303 90.8% 211.110:00 1,330 1,286 3.3% 22 307 98.4% 223.110:30 1,364 1,338 1.9% 33 296 99.0% 222.511:00 1,380 1,280 7.2% 34 306 98.2% 222.011:30 1,272 1,247 2.0% 44 298 94.6% 218.012:00 1,179 1,177 0.2% 1 306 91.6% 218.312:30 1,174 1,160 1.2% 10 302 95.5% 203.813:00 1,018 999 1.9% 9 314 95.4% 182.913:30 1,061 961 9.4% 67 306 100.0% 163.414:00 1,173 1,082 7.8% 78 313 99.5% 188.914:30 1,212 1,179 2.7% 23 304 96.6% 206.115:00 1,137 1,122 1.3% 15 320 96.9% 205.815:30 1,169 1,137 2.7% 17 311 97.1% 202.216:00 1,107 1,059 4.3% 46 315 99.2% 187.116:30 914 892 2.4% 22 307 95.2% 160.017:00 615 615 0.0% 2 328 83.0% 135.017:30 420 420 0.0% 0 328 73.8% 103.518:00 49 49 0.0% 14 180 84.2% 5.8
7
Beyond Averages: Waiting Times in a Call Center
Small Israeli Bank
quantiles of waiting times to those of the exponential (the straight line at the right plot). The �t is reasonableup to about 700 seconds. (The p-value for the Kolmogorov-Smirnov test for Exponentiality is however 0 {not that surprising in view of the sample size of 263,007).
Figure 9: Distribution of waiting time (1999)
Time
0 30 60 90 120 150 180 210 240 270 300
29.1 %
20 %
13.4 %
8.8 %
6.9 %5.4 %
3.9 %3.1 %
2.3 % 1.7 %
Mean = 98SD = 105
Waiting time given agent
Exp
qua
ntile
s
0 200 400 600
020
040
060
0
Remark on mixtures of independent exponentials: Interestingly, the means and standard deviations in Table19 are rather close, both annually and across all months. This suggests also an exponential distributionfor each month separately, as was indeed veri�ed, and which is apparently inconsistent with the observerdannual exponentiality. The phenomenon recurs later as well, hence an explanation is in order. We shall besatis�ed with demonstrating that a true mixture W of independent random varibles Wi, all of which havecoeÆcients of variation C(Wi) = 1, can also have C(W ) � 1. To this end, let Wi denote the waiting time inmonth i, and suppose it is exponentially distributed with meanmi. Assume that the months are independentand let pi be the fraction of calls performed in month i (out of the yearly total). If W denotes the mixtureof these exponentials (W =Wi with probability pi, that is W has a hyper-exponential distribution), then
C2(W ) = 1 + 2C2(M);
where M stands for a �ctitious random variable, de�ned to be equal mi with probability pi. One concludesthat if themi's do not vary much relative to their mean (C(M) << 1), which is the case here, then C(W ) � 1,allowing for approximate exponentiality of both the mixture and its constituents.
6.2.1 The various waiting times, and their rami�cations
We �rst distinguished between queueing time and waiting time. The latter does not account for zero-waits,and it is more relevant for managers, especially when considered jointly with the fraction of customers thatdid wait. A more fundamental distinction is between the waiting times of customer that got served and thosethat abandoned. Here is it important to recognize that the latter does not describe customers' patience,which we now explain.
A third distinction is between the time that a customer needs to wait before reaching an agent vs. the timethat a customer is willing to wait before abandoning the system. The former is referred to as virtual waitingtime, since it amounts to the time that a (virtual) customer, equipped with an in�nite patience, would havewaited till being served; the latter will serve as our operational measure of customers' patience. While bothmeasures are obviously of great importance, note however that neither is directly observable, and hence mustbe estimated.
25
Large U.S. BankChart1
Page 1
0
2
4
6
8
10
12
14
16
18
20
2 5 8 11 14 17 20 23 26 29 32 35
Time
Relative frequencies, %
Medium Israeli Bankwaitwait
Page 1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380
Time (Resolution 1 sec.)
Relative frequencies, %
8
The Second Prerequisite: Models
Through Examples Only.
Each example starts with a Complex Reality and ends with a usefulinsight due to a Simple Model.
‘‘Theorem": A useful model must be simple (yet not too simple).
Models in decreasing simplicity-levels:
I Conceptual: Service Networks = Queueing Networks
I Descriptive: Averages, Histograms
I Explanatory: Comparative, Regression
I Analytical/Mathematical: Little’s Law, Fluid Models, QueueingModels, Diffusion Refinements.
“Corollary": To be useful, a simple model sometimes requires deepanalysis.
9
Conceptual Model: Face-to-Face Services
Bank Branch = Queueing Network
23
Teller
Entrance
Tourism
Xerox
Manager
Teller
Entrance
Tourism
Xerox
Manager
Bottleneck!
10
Descriptive Model: Transition Probabilities (Averages)
Bank: A Queuing Network
Transition Frequencies Between Units in The Private and Business Sections:
Private Banking Business
To Unit Bankers Authorized Compens - Tellers Tellers Overdrafts Authorized Full Exit
From Unit Personal - ations Personal Service
Bankers 1% 1% 4% 4% 0% 0% 0% 90%
Private Authorized Personal 12% 5% 4% 6% 0% 0% 0% 73%
Banking Compensations 7% 4% 18% 6% 0% 0% 1% 64%
Tellers 6% 0% 1% 1% 0% 0% 0% 90%
Tellers 1% 0% 0% 0% 1% 0% 2% 94%
Services Overdrafts 2% 0% 1% 1% 19% 5% 8% 64%
Authorized Personal 2% 1% 0% 1% 11% 5% 11% 69%
Full Service 1% 0% 0% 0% 8% 1% 2% 88%
Entrance 13% 0% 3% 10% 58% 2% 0% 14% 0%
Legend: 0%-5% 5%-10% 10%-15% >15%
Dominant Paths - Business:
Unit Station 1 Station 2 Total Parameter Tourism Teller Dominant Path
Service Time 12.7 4.8 17.5 Waiting Time 8.2 6.9 15.1
Total Time 20.9 11.7 32.6
Service Index 0.61 0.41 0.53
Dominant Paths - Private:
Unit Station 1 Station 2 Total Parameter Banker Teller Dominant Path
Service Time 12.1 3.9 16.0 Waiting Time 6.5 5.7 12.2
Total Time 18.6 9.6 28.2
Service Index 0.65 0.40 0.56
Service Index = % time being served
11
Conceptual Model: Hospital (ED) Network(Marmur, Sinreich)
PhysicianNurseImagingLabElse
60estimated max time
initialexamination
decision point for alternative processes
10%probability of events
06vital signs
07
E.C.G05
decision
awaitingdischarge
40
treatment 41
50
consultation
instructionsprior discharge
discharge /hospitalization
else
triage04
43
54
reception03
observation
46
every 15 minutes
follow up47
bloodwork
1312
100%
imaging /consultation /treatment
17
14
decision
20
ultrasound2928
21
Xray
2725,26
CT
3130
22
15
39
37
45
follow up48
every 15 minutes
49
11
handlingpatient&family
08
09
38imaging
36
3534
32,33 treatment18
56
hospitalization/discharge
awaitingevacuation
55
52
53
10
treatment 19
16
discharge
51
else
treatment42
44
reference point
labs labs
consultation
Labs
consultation
imaging
decision
proportion of patients 01 process requires bed 02
23
24
Figure 2. The Unified Patient Process Chart
12
Descriptive Model: Service Times (Averages) or,Even “Doctors" Can Manage
Afternoon, by Case
Morning, by Hour
Operations Time In a Hospital
Operations Time Histogram: Operations Time - Morning (by Hour) vs. Afternoon (by Case): Ethical?
Even Doctors Can Manage!
0%2%4%6%8%
10%12%14%16%18%20%
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14
Hours
Freq
uenc
y
AVG: 2.08 HoursSTD: 4.12 HoursSample Size: 4347
CV >> 1
0
1
2
3
4
5
6
EEG Orthopedics Surgery Blood Surgery Plastic Surgery Heart/ChestSurgery
Neuro-Surgery Eyes E.I. Surgery
Department
Hou
rs
AM
PM
20
13
Conceptual Model: The “Production of Justice"
The Labor-Court Process in Haifa, Israel
“Production” Of Justice
Queue
Mile Stone
Activity
Appeal
Proceedings
Closure
Prepare AllocateOpen File
Avg. sojourn time ≈ in months / years
Processing time ≈ in mins / hours / days
Phase Transition
Phase
14
Analytical Model: Little’s Law in Court (still Averages)
Operational Performance: Rate λ & Time W
45 100
118
59
33
.
.
.
..
(6.2, 7.4) (13.5, 7.4)
(26.3, 4.5)
(12, 4.9)
(7.2, 4.6) 3
001
3
001
01
0
3
01
3
00
3
01
0
1
2
3
4
5
6
7
8
9
10
0 5 10 15 20 25 30
Judges: Performance Analysis
Case Type 0 Judge1 Case Type 01 Judge2 Case Type 3 Judge3 Judge4 Judge5
Judges: Performance Analysis Judges: Performance by Case-Type
Ave
rage
Num
ber o
f Mon
ths -
W
Average Number of Cases / Month - λ
Judges: Operational Performance – Base Case
15
Expertise “invite" Skills-Base-Routing (SBR)
Operational Performance: Judges’ Heterogeneity (Best / Worse)
45 100
118
59
33
.
.
.
..
(6.2, 7.4) (13.5, 7.4)
(26.3, 4.5)
(12, 4.9)
(7.2, 4.6) 3
001
3
001
01
0
3
01
3
00
3
01
0
1
2
3
4
5
6
7
8
9
10
0 5 10 15 20 25 30
Judges: Performance Analysis
Case Type 0 Judge1 Case Type 01 Judge2 Case Type 3 Judge3 Judge4 Judge5
Ave
rage
Num
ber o
f Mon
ths -
W
Average Number of Cases / Month - λ
Judges: Operational Performance – Base Case Judges: Performance Analysis Judges: Performance by Case-Type
16
Analytical Model: Little’s Law in Court (still Averages)
Judges’ Profiles (Operational)
3
001
3
001
01
0
3
01
3
00
3
01
0
1
2
3
4
5
6
7
8
9
10
0 5 10 15 20 25 30
(6.2, 7.4) (13.5, 7.4)
(26.3, 4.5)
(12, 4.9)
(7.2, 4.6)
.
.
.
..
Judges: Performance Analysis
Case Type 0 Judge1 Case Type 01 Judge2 Case Type 3 Judge3 Judge4 Judge5
Avg
. Mon
ths -
W
Avg. Cases / Month - λ
17
Analytical Model: Little’s Law in Court (still Averages)
Judges: The Best/Worst (Operational) Performer
(6.2, 7.4) (13.5, 7.4)
(26.3, 4.5)
(12, 4.9)
(7.2, 4.6) 3
001
3
001
01
0
3
01
3
00
3
01
0
1
2
3
4
5
6
7
8
9
10
0 5 10 15 20 25 30
.
.
.
..
45 100
118
59
33
Judges: Performance Analysis
Case Type 0 Judge1 Case Type 01 Judge2 Case Type 3 Judge3 Judge4 Judge5
Avg
. Mon
ths -
W
Avg. Cases / Month - λ
18
Call-Center Network: Gallery of Models
Agents(CSRs)
Back-Office
Experts)(Consultants
VIP)Training (
Arrivals(Business Frontier
of the21th Century)
Redial(Retrial)
Busy)Rare(
Goodor
Bad
Positive: Repeat BusinessNegative: New Complaint
Lost Calls
Abandonment
Agents
ServiceCompletion
Service Engineering: Multi-Disciplinary Process View
ForecastingStatistics
New Services Design (R&D)Operations,Marketing
Organization Design:Parallel (Flat)Sequential (Hierarchical)Sociology/Psychology,Operations Research
Human Resource Management
Service Process Design
To Avoid Delay
To Avoid Starvation Skill Based Routing
(SBR) DesignMarketing,Human Resources,Operations Research,MIS
Customers Interface Design
Computer-Telephony Integration - CTIMIS/CS
Marketing
Operations/BusinessProcessArchiveDatabaseDesignData Mining:MIS, Statistics, Operations Research, Marketing
InternetChatEmailFax
Lost Calls
Service Completion)75% in Banks (
( Waiting TimeReturn Time)
Logistics
Customers Segmentation -CRM
Psychology, Operations Research,Marketing
Expect 3 minWilling 8 minPerceive 15 min
PsychologicalProcessArchive
Psychology,Statistics
Training, IncentivesJob Enrichment
Marketing,Operations Research
Human Factors Engineering
VRU/IVR
Queue)Invisible (
VIP Queue
(If Required 15 min,then Waited 8 min)(If Required 6 min, then Waited 8 min)
Information DesignFunctionScientific DisciplineMulti-Disciplinary
IndexCall Center Design
(Turnover up to 200% per Year)(Sweat Shops
of the21th Century)
Tele-StressPsychology
19
The “Phases of Waiting" for Service
Common Experience:I Expected to wait 5 minutes, Required to 10I Felt like 20, Actually waited 10 (hence Willing ≥ 10)
An attempt at “Modeling the Experience":1. Time that a customer expects to wait2. willing to wait ((Im)Patience: τ )3. required to wait (Offered Wait:V )4. actually waits (Wq = min(τ, V ))5. perceives waiting.
Experienced customers ⇒ Expected = Required“Rational" customers ⇒ Perceived = Actual.
Then left with (τ, V ).
20
Call Center Data: Hazard Rates (Un-Censored)
Israel
U.S.
(Im)Patience Time τ
0 50 100 150 2000
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5x 10
−3
time, sec
haza
rd r
ate
0 10 20 30 40 50 600
0.05
0.1
0.15
0.2
0.25
0.3
0.35
time, sec
haza
rd r
ate
Required/Offered Wait V
36
0 10 20 30 40 50 600
2
4
6
8
10
12
14
16
time, sec
haza
rd r
ate
actuarial estimatespline smoother
21
A Patience Index
Quantifying (Im)Patience: “Willing to wait 15 min" = Patient?
Theoretical Patience-Index ∆=
Willing to waitExpected to wait
=E[τ ]
E[V ],
“assuming" Experienced; further “assuming" that τ and V areExponential, the M-L estimate of Index is the easily-measurable:
Empirical Patience-Index ∆=
% Served% Abandoning
Index Validation: Theoretical vs. Empirical
0
1
2
3
4
5
6
7
8
9
10
2 3 4 5 6 7 8 9
Empirical Index
Theo
retic
al In
dex
22
Predicting Performance
Model Primitives:
I Arrivals to service (random process)I (Im)Patience while waiting τ (r.v.)I Service times (r.v.)I # Servers / Agents (parameter / r.v.)
Model Output: Offered-Wait V (r.v.)
Operational Performance Measure calculable in terms of (τ, V ).
I eg. Average Wait = E[min{τ, V}]I eg. % Abandonment = P{τ < V}
Application: Staffing – How Many Agents? (When? Who?)
23
The Basic Staffing Model: Erlang-A (M/M/n +M)
agents
arrivals
abandonment
λ
µ
1
2
n
…
queue
θ
Erlang-A Parameters:
I λ – Arrival rate (Poisson)I µ – Service rate (Exponential)I θ – Impatience rate (Exponential)I n – Number of Service-Agents.
24
Erlang-A: Fitting a Simple Model to a Complex Reality
Hourly Performance vs. Erlang-A Predictions (1 year)
% Abandon E[Wait] %{Wait > 0}
0 0.1 0.2 0.3 0.4 0.5 0.60
0.1
0.2
0.3
0.4
0.5
Probability to abandon (Erlang−A)
Pro
babi
lity
to a
band
on (
data
)
0 50 100 150 200 2500
50
100
150
200
250
Waiting time (Erlang−A), sec
Wai
ting
time
(dat
a), s
ec
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Probability of wait (Erlang−A)
Pro
babi
lity
of w
ait (
data
)
I Empirically-Based & Theoretically-Supported Estimation of(Im)Patience: θ̂ = P{Ab}/E[Wq])
I Small Israeli Bank (more examples in progress)
25
Testing the Erlang-A Primitives
I Arrivals: Poisson?I Service-durations: Exponential?I (Im)Patience: Exponential?
Validation: Support? Refute?
26
Arrivals to Service: only Poisson-Relatives
Arrival Rate to Three Call Centers
Dec. 1995 (U.S. 700 Helpdesks) May 1959 (England)
Q-Science
May 1959!
Dec 1995!
(Help Desk Institute)
Arrival Rate
Time 24 hrs
Time 24 hrs
% Arrivals
(Lee A.M., Applied Q-Th)
28
Q-Science
May 1959!
Dec 1995!
(Help Desk Institute)
Arrival Rate
Time 24 hrs
Time 24 hrs
% Arrivals
(Lee A.M., Applied Q-Th)
28
November 1999 (Israel)
Arrival Process, in 1999
Yearly Monthly
Daily Hourly
Observation:Peak Loads at 10:00 & 15:00
27
Service Durations: LogNormal Prevalent
Israeli Bank Survival-FunctionsLog-Histogram by Service-Class
0
100
200
300
400
500
600
700
800
900
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8
Log(service time)
Freq
uenc
y
frequency normal curve
Average = 2.24St.dev. = 0.42
31
Service TimeSurvival curve, by Types
Time
Surv
ival
Means (In Seconds)
NW (New) = 111
PS (Regular) = 181
NE (Stocks) = 269
IN (Internet) = 381
34
I New Customers: 2 min (NW);I Regulars: 3 min (PS);
I Stock: 4.5 min (NE);I Tech-Support: 6.5 min (IN).
Observation: VIP require longer service times.
28
(Im)Patience while Waiting (Palm 1943-53)
Irritation ∝ Hazard Rate of (Im)Patience DistributionRegular over VIP Customers – Israeli Bank
14
16
I Peaks of abandonment at times of AnnouncementsI Call-by-Call Data (DataMOCCA) required (& Un-Censoring).
Observation: VIP are more patient (Needy)
29
A “Service-Time" Puzzle at a Small Israeli BankInter-related Primitives
Average Service Time over the Day – Israeli Bank
�
�
�
�
Figure 12: Mean Service Time (Regular) vs. Time-of-day (95% CI) (n =
42613)
Time of Day
Mea
n S
ervi
ce T
ime
10 15 20
100
120
140
160
180
200
220
240
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
3011
Prevalent: Longest services at peak-loads (10:00, 15:00). Why?Explanations:
I Common: Service protocol different (longer) during peak times.I Operational: The needy abandon less during peak times;
hence the VIP remain on line, with their long service times.30
Erlang-A: Simple, but Not Too Simple
Experience:I Arrival process not pure Poisson (time-varying, σ2 too large)I Service times not exponential (typically close to lognormal)I Patience times not exponential (various patterns observed).I Customers and Servers not homogeneous (classes, skills)
Questions naturally arise:
1. Why does Erlang-A practically work? justify robustness.2. When does it fail? chart boundaries.3. Generalize: time-variation, SBR, networks, uncertainty , . . .
Answers via Asymptotic Analysis, as load- and staffing-levels ↑ ,which captures model-essentials:
I Efficiency-Driven (ED) regime: Fluid models (deterministic).I Quality- and Efficiency-Driven (QED) regime: Diffusion
refinements (eg. revealing that the patience-density at the originis “all" that is needed).
31
DataMOCCA = Data MOdels for Call Center Analysis
I Technion: P. Feigin, V. Trofimov, Statistics / SEE Laboratory.I Wharton: L. Brown, N. Gans, H. Shen (UNC).I industry:
I U.S. Bank: 2.5 years, 220M calls, 40M by 1000 agents.I Israeli Cellular: 2.5 years, 110M calls, 25M calls by 750 agents;
ongoing.
Project Goal: Designing and Implementing a (universal)data-base/data-repository and interface for storing, retrieving,analyzing and displaying Call-by-Call-based Data / Information.
System Components:I Clean Databases: operational-data of individual calls / agents.I Graphical Online Interface: easily generates graphs and tables,
at varying resolutions (seconds, minutes, hours, days, months).
Free for academic adoption: Mini version available on a DVD;working version 7GB tables, or 20GB raw zipped, for each callcenter – ask for my mini-HD.
32
Arrivals to Service: Predictable vs. RandomArrival Rates on Tuesdays in a September – U.S. Bank
0
500
1000
1500
2000
2500
3000
3500
0:00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00
Time (Resolution 30 min.)
Arrival rate
04.09.2001 11.09.2001 18.09.2001 25.09.2001
I Tuesday, September 4th: Heavy, following Labor Day.I Tuesdays, September 18 & 25: Normal.I Tuesday, September 11th, 2001.
33
A “Waiting-Times" Puzzle at a Medium Israeli Bankwaitwait
Page 1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380
Time (Resolution 1 sec.)Relative frequencies, %
Peaks Every 60 Seconds. Why?I Human: Voice-announcement every 60 seconds.I System: Priority-upgrade (unrevealed) every 60 seconds.
Served Customers Abandoning Customerswaithandled
Page 1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380
Time (Resolution 1 sec.)
Relative frequencies, %
waitab
Page 1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340
Time (Resolution 1 sec.)
Relative frequencies, %
34
Priorities, Economies-of-Scale, SBR
Regular vs. VIP Customers: Cellular – March 23, 2004
Average Wait Staffing Level
0
5
10
15
20
25
30
35
40
45
50
7:00 9:00 11:00 13:00 15:00 17:00 19:00 21:00 23:00
Time
Average waiting time, sec
Private Private Platinum
0
10
20
30
40
50
60
7:00 9:00 11:00 13:00 15:00 17:00 19:00 21:00 23:00
Time
Number of agents
Private Private Platinum
I Design: VIP-dedicated agents, Regular-dedicated Agents.I VIP’s are not served better than Regular’s
I Solutions: Add VIP agents (costly), or Re-Design.
35
Priorities and Routing Protocols I
Regular vs. VIP Customers: Cellular – October 2004
Delay Probability Average Wait
0
10
20
30
40
50
60
70
80
90
100
7:00 9:00 11:00 13:00 15:00 17:00 19:00 21:00 23:00
Time (Resolution 30 min.)
Delay Probability, %
Private Private Platinum
0
5
10
15
20
25
30
35
40
7:00 9:00 11:00 13:00 15:00 17:00 19:00 21:00 23:00
Time (Resolution 30 min.)
Average Wait, sec
Private Private Platinum
More VIPs delayed than Regulars, yet their average wait is shorter.
What changed since last March?
36
Priorities and Routing Protocols II
Waiting-Time Histograms: Cellular – October 2004
Regular Customers VIP (Platinum) Customersprivate_hist
Page 1
0.0
0.5
1.0
1.5
2.0
2.5
3.0
2 14 26 38 50 62 74 86 98
Time (Resolution 1 sec.)
Relative frequencies, %
platinum_hist
Page 1
0.0
2.0
4.0
6.0
8.0
10.0
12.0
14.0
16.0
18.0
2 14 26 38 50 62 74 86
Time (Resolution 1 sec.)
Relative frequencies, %
After 25 seconds of wait, VIP customers are routed with highpriority to Regular agents. Hence, almost no long waiting times forVIP’s.
37
Network Balancing via Inter-Queues at a U.S. Bank
73
Distributed Call Center (U.S. Bank)
NY 1
RI 3
PA 2
MA 4
179+5
619+3
11+1
74+7
8+1
19+1
20
508+2
101+2
2
External arrivals:2092 2063(98.6%Served)+29(1.4%Aban)
Not Interqueued:1209(57.8%)
• Served: 1184(97.9/56.6)
• Aban: 25(2.1/1.2) Interqueued :883(42.2) • Served
here:174(19.7/8.3)
• Served at 2: 438(49.6/20.9) S d t 3
External arrivals: 16941687(99.6%
Served)+7( 0.4% Aban)
Not Interqueued: 1665(98.3)
• Served: 1659 (99.6/97.9)
• Aban: 6 (0.4/04) Interqueued:28+1 (1.7)
• Served here: 17(58.6/1)
• Served at 1: 3(10.3/0.2)
External arrivals: 122 112(91.8
Served)+10(8.2 Aban)
Not Interqueued: 93 (76.2)
• Served: 85 (91.4/69.7)
• Aban: 8 (8.6/6.6) Interqueued:27+2
(23.8) • Served here:
14(48.3/11.5) • Served at 1: 6
External arrivals: 1770 1755(99.2
Served)+15(0.8 Aban)
Not Interqueued: 1503(84.9)
• Served: 1497 (99.6/84.6)
• Aban: 6 (0.4/0.3) Interqueued:258+9
(15.1) • Served here: 110
(41.2/6.2) • Served at 1:58
(21 7/3 3)
Internal arrivals: 224
• Served at 1: 67 (29.9)
• Served at 2: 41 (18.3)
• Served at 3: 87 (38.8)
• Served at 4:
Internal arrivals: 643
• Served at 1: 157 (24.4)
• Served at 2: 195 (30.3)
• Served at 3: 282 (43.9)
• Served at 4: 4 (0.6)
• Aban at 1: 3
Internal arrivals: 81
• Served at 1: 17(21)
• Served at 3: 42(51.9)
• Served at 4: 15(18 5)
Internal arrivals: 613• Served at 1:
41(6.7) • Served at 2:
513(83.7) • Served at 3:
55(9.0) • Aban at 1:
2(0.3)
10 AM – 11 AM (03/19/01): Interflow Chart Among the 4 Call C t f Fl t B k
38
Balancing Protocols and Performance Level
U.S. Bank: Histograms of Waiting Times
Retail BusinessChart1
Page 1
0
2
4
6
8
10
12
14
16
18
20
2 5 8 11 14 17 20 23 26 29 32 35
Time
Relative frequencies, %
Sheet3 Chart 1
Page 1
0
2
4
6
8
10
12
2 8 14 20 26 32 38 44 50 56 62 68 74 80 86 92
Time
Relative frequencies, %
Peak for Retail service at 10 seconds – Why?After 10 seconds of wait, Retail customers sent into the inter-queue.
Business customers – peak at 5 seconds, for the same reason.Second peak – unclear, maybe priority-upgrade.
39
Data-Based Service-Research(with DataMOCCA, even before tenure)
I Contrast with EmpOM: Industry / Company / Survey Data(Social Sciences)
I Converge to: Measure, Model, Validate, Experiment, Refine(Physics, Biology, . . .).
I Prerequisites:I OR, OM, IE, (Mktg.) - for DesignI CS, IS, Stat. - for Implementation.
I Outcomes: Relevance, Credibility; Interest, Fun;Call Centers as a Pilot (eg. for Healthcare). Moreover,
I Teaching: Class, Homework (Experimental Data Analysis); Cases.I Research: Validate Existing (Queueing) Theory/Laws and Suggest
New Models/Research.I Practice: OM Tools (Scenario Analysis), Mktg. (Trends,
Benchmarking).
40
Live Demonstration of DataMOCCA5-7 minutes, to emphasize “online" capabilities.
U.S. Bank
I Daily Reports: October 2003, weekdays; typically takes 10-20sec till a first output, but this is because of PowerPoint/Windows.Then do few additional Daily Reports, say Monday, Tuesday,...(starting with STATCCA, as opposed to by minimizing thepowerpoint screnn) - this will be now happening very fast.
I Time-Series: Number of agents, for ALL classes, all months,weekdays. (Including total). Shows scale, trends. Then doService Durations, indicating that 1 second of 1000 agents couldcost $500M per year. Could also do Unhandled (lower middleentry in list), for only Retail and Premium - Premium is worse,and deteriorating,
I Daily Summaries:I Tuesdays in September 2001; September 11th; shown during
lecture under the heading “Predictable or Random";I 30 sec scale - stoch. variability, 1 hour scale - the “right" scale;I % to mean, to show very similar shape over 3 Tuesdays. Suggests
the model λ(t) = λ0(t) · Z , for 0 ≤ t ≤ T .41