Managing Call Centers with Many Strategic Agents

Transcription

1 Managing Call Centers with Many Strategic Agents Rouba Ibrahim, Kenan Arifoglu Management Science and Innovation, UCL YEQT Workshop - Eindhoven November 2014

2 Work Schedule Flexibility 86SwofwthewqBestwCompanieswtowWorkwForqwofferw employeeswsomewtypewofwflexiblewschedule. -3Fortune, Swofwemployerswallowedwsomewworkerswtow periodicallywchangewtheirwstartingwandwquittingwtimes.w -3National3Study3of3Employers3(USA), InwthewUS,wallwfederalwemployeeswnowwhavewthew rightwtowrequestwmorewflexiblewworkwoptions.ww -ywhiteyhouseymemoyissuedyonyjuney23,y2014.y InwthewUK,wthewrightwtowrequestwflexiblewworkingw hourswiswnowwextendedwtowallwworkers.ww -yflexibleyworkingylawyissuedyonyjulyy1,y2014.y

3 Flexitime at Hydro-Québec Flexitime schedule: 30% of call center agents have flexible work hours Must be there during core periods (2 periods of 2 hours each) Are otherwise free to choose their time schedules Don t need to inform their managers ahead of time Must be there for a total of 70 hours per 2 weeks

4 Flexitime at Hydro-Québec Flexitime schedule: 30% of call center agents have flexible work hours Must be there during core periods (2 periods of 2 hours each) Are otherwise free to choose their time schedules Don t need to inform their managers ahead of time Must be there for a total of 70 hours per 2 weeks Major headaches for Hydro-Québec s call center managers!

5 Research Questions What is the impact of strategic agent behavior on the system? What are optimal operational decisions with strategic agents? Can we align the objectives of the system manager and the agents?

6 Related Literature Staffing and routing in many-server systems Queueing games Strategic servers in queues Gopalakrishnan, R., Dorouli, S., Ward, A. and A. Weirman Routing and staffing when servers are strategic. Gurvich, I., Lariviere, M. and A. Moreno-Garcia Staffing service systems when capacity has a mind of its own.

7 Staffing Decisions with Non-Strategic Agents

8 Non-Strategic Agents Queueing Framework µ µ θ θ λ A Period A λ B Period B Poisson arrival processes with rates λ A, λ B No overlap between periods A and B I.I.D. exponential service times with rate µ = 1 I.I.D. exponential abandonment times with rate θ

9 Non-Strategic Agents System Manager s Problem with Non-Strategic Agents Staffing Costs c: Fixed compensation r: Variable compensation rate t A, t B : Total compensations t A = c + r λa n A and t B = c + r λb n B. Customer-Related Costs p: Abandonment penalty cost ($ per customer who abandons) h: Delay cost ($ per customer per minute wait)

10 Non-Strategic Agents Fluid Approximation to the System Manager s Problem For t = A, B: n t : Number of servers in period t q t : Queue length in period t η t : Total abandonment rate in period t η t = (λ t n t ) + and q t = η t θ. The system manager determines optimal staffing levels n A and n B : ( min {n A n A 0,n B 0 + c + r λ A n A ( p + h θ ) ( + n B c + r λ B n B ) (λ A n A ) + + ) ( p + h θ ) (λ B n B ) + } Bassamboo and Randhawa (2010).

11 Non-Strategic Agents Fluid-Based Prescriptions with Non-Strategic Agents Recall: c = staffing cost; p = abandonment cost; h = delay cost. Optimal Staffing Levels (i) if c p + h/θ, then n A = λ A and n B = λ B, (ii) if c > p + h/θ, then n A = n B = 0. In (i), both A and B are critically loaded. In this case, n A α = na + n B = λ A λ A + λ B and ρ A = ρ B = 1, where ρ t = λ t /n t.

12 Strategic Agents Staffing Decisions with Strategic Agents

13 Strategic Agents Strategic Agent Behavior µ µ θ θ λ A Period A λ B Period B Agents have individual valuations for working in either period Strategic agents select working period of their choice

14 Strategic Agents A Two-Stage Sequential Game Sequence of events Stage 1: System manager selects total staffing level n Stage 2: Agents in pool of size n select working period An equilibrium arises in the system. We determine this equilibrium by using backward induction.

15 Strategic Agents Second Stage: Agent s Problem

16 Strategic Agents Individual Agent Preferences n: Total number of agents in the system (given) n A, n B : To be determined by the agents Independent and heterogeneous agents v A, v B : agent valuations v A (n A ) = c + r λa n A and v B (n B ) = c + (r + X ) λb n B, where X is a random variable (cdf F, ccdf F ). Agent selects period A if, and only if, v A (n A ) v B (n B ).

17 Strategic Agents Subgame Perfect Nash Equilibrium Let α e be the equilibrium proportion of agents who select A. Existence and Uniqueness of Equilibrium We show that there exists a unique equilibrium threshold, t e, such that α e = P(X t e ) = F (t e ). We also show that t e is the unique solution of: r λ A F (t e ) λ B F (t e ) r te = 0.

18 Strategic Agents Impact of Strategic Agent Behavior Recall that we have at optimum in the non-strategic case: α = λ A /(λ A + λ B ) is the optimal proportion of agents who are in A ρ A = ρ B = 1 Comparison with the Non-Strategic Case We show that, for any value of n, the following hold: (i) F (0) = α α e = α (ii) F (0) < α α e < α (iii) F (0) > α α e > α where F (0) = P(X 0).

19 Strategic Agents Explanation: Case (i) Suppose that the staffing levels in periods A and B are as follows: Period A Period B n A = λ A Then, let agents select their periods. n B = λ B The proportion of agents who choose period A is: ( P(v B v A ) = P c + (r + X ) λ B c + r λ ) A = P(X 0) = F (0). n B n A So, for the above to be an equilibrium state, we need to have: F (0) = In this case, α e = F (0) = α. λ A λ A + λ B = α.

20 Strategic Agents Case (ii): System Imbalance Now, what happens if F (0) < α? Period A Period B n A < λ A n B > λ B The proportion of agents who choose period A is: ( P(v B v A ) = P c + (r + X ) λ B c + r λ ) A = P(X 0) = F (0) < α. n B n A In this case, some agents in A will now select B instead. So, our initial state cannot be an equilibrium, and we have: F (0) < α e < α. In this case, period A becomes overloaded and period B underloaded.

21 Strategic Agents First Stage: System Manager s Problem

22 Strategic Agents Fluid Approximation with Strategic Agents In this stage, assume that α e is given. The system manager determines the optimal total staffing level n: ( min {nα e n 0 + c + r λ A nα ( e p + h θ ) ( + n(1 α e ) ) (λ A nα e ) + + ) λ B c + r n(1 α e ) ( p + h ) } (λ B n(1 α e )) + θ

23 Strategic Agents Fluid-Based Prescriptions: Asymptotically Optimal Regimes It is never optimal to have: Both A and B underloaded Both A and B overloaded So, have at least one period critically loaded. A overloaded B critically loaded A critically loaded B underloaded A underloaded B critically loaded A critically loaded B overloaded 0 c/(p+h/θ) α * 1 c/(p+h/θ) 1 α e

24 Strategic Agents Cost Impact of Strategic Agents Percent cost increase α * 10 0 c = 0.5, p = 0.7, h = 1, and µ = θ = 1 λ A = 150 and λ B = 250 X uniform on (x l, x u ) Vary r, x l, and x u to vary α e α e

25 System Controls

26 A Coordinating Compensation Scheme Let r A = r + β A and r B = r + β B be the variable compensation rates for periods A and B. Unequal Compensation Rates We show that the compensation scheme with β A = (1 α )F 1 (α ) and β B = α F 1 (α ), is a budget-neutral compensation scheme for which α e (r A, r B ) = α. The system manager can control the system by increasing compensation in the period which is busier at equilibrium.

27 Conclusions We proposed a model for strategic agents We quantified the impact of strategic agent behavior We derived optimal staffing levels with strategic agents We developed a compensation scheme that makes agents behave in line with the system manager s objective

28 Extensions We also considered the following extensions: Social welfare Different agent valuation models Routing Decisions What if the system manager doesn t honor or only partially honors agents choices? General abandonment Heterogeneous service rates

29 Thank You!