Ethically Aligned Opportunistic Scheduling for Productive Laziness
Han Yu, Chunyan Miao, Yongqing Zheng, Lizhen Cui, Simon Fauvel, Cyril Leung
aa r X i v : . [ c s . A I] J a n Ethically Aligned Opportunistic Scheduling for Productive Laziness
Han Yu , , , Chunyan Miao , , , Yongqing Zheng , , Lizhen Cui , Simon Fauvel and Cyril Leung School of Computer Science and Engineering, Nanyang Technological University (NTU), Singapore Joint NTU-UBC Research Centre of Excellence in Active Living for the Elderly (LILY), NTU, Singapore Alibaba-NTU Singapore Joint Research Institute, Singapore School of Software Engineering, Shandong University, Jinan, Shandong, China Shanda Dareway Software Co. Ltd, Jinan, Shandong, China Department of Electrical and Computer Engineering, The University of British Columbia, Vancouver, BC, CanadaCorresponding Authors: [email protected], [email protected]
Abstract
In artificial intelligence (AI) mediated workforce manage-ment systems (e.g., crowdsourcing), long-term success de-pends on workers accomplishing tasks productively and rest-ing well. This dual objective can be summarized by the con-cept of productive laziness. Existing scheduling approachesmostly focus on efficiency but overlook worker wellbeingthrough proper rest. In order to enable workforce manage-ment systems to follow the IEEE Ethically Aligned Designguidelines to prioritize worker wellbeing, we propose a dis-tributed
Computational Productive Laziness (CPL) approachin this paper. It intelligently recommends personalized work-rest schedules based on local data concerning a worker’s ca-pabilities and situational factors to incorporate opportunisticresting and achieve superlinear collective productivity with-out the need for explicit coordination messages. Extensive ex-periments based on a real-world dataset of over 5,000 work-ers demonstrate that CPL enables workers to spend 70% ofthe effort to complete 90% of the tasks on average, providingmore ethically aligned scheduling than existing approaches.
Introduction
In today’s world, artificial intelligence (AI) hasbeen employed to manage large-scale workforcessuch as crowdsourcing systems (Miao et al. 2016;Michelucci and Dickinson 2016; Pan et al. 2016). Forexample, in DiDi Chuxing, AI technologies are deployedto dynamically match drivers to tasks in order to enhanceoperation efficiency (Xu et al. 2018). Human workersbecome fatigued or bored over long sessions of work,which can cause inefficacy (Leiter and Maslach 2015). Arecent study found that short breaks significantly improveworkers’ motivation while maintaining the quality of work(Dai et al. 2015). However, existing AI approaches forworkforce management in crowdsourcing mostly do notexplicitly incorporate rest into their recommendations(Chai et al. 2017).From an ethically aligned design perspective(The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems 2018),it is desirable to incorporate breaks into scheduling ap-proaches in order to prioritize workers’ wellbeing. Never-theless, as highlighted in (Yu et al. 2018), ethically aligned
Copyright c (cid:13)
AI approaches need to be designed such that they balancethe concern for human wellbeing while still achieve theirdesign objectives. In this paper, we set out to address thischallenge by proposing an ethically aligned opportunisticscheduling approach that can achieve “productive laziness”-
Computational Productive Laziness (CPL).Originally conceptualized in social sciences literature,productive laziness (Whillans et al. 2017) is a rule-of-thumbguideline on how workers should approach their work inorder to achieve work-life balance. The general idea is towork when you are highly efficient, and rest otherwise. CPLcoordinates workers to work when situational factors in-duce high efficiency or demand working, and rest whenthey do not. Thus, it is important to identify factors influ-encing workers’ productivity and the urgency of work. Weformulate the problem of achieving productive laziness inlarge-scale systems as a multi-objective constrained opti-mization problem. Following the Lyapunov optimization-based framework (Yu et al. 2016), we analyse the interactiondynamics involved in a workforce management system andderive a novel
Work-Rest Index (WRI) which expresses theinterplay among four aspects affecting worker effort output:1.
Situational factors : the current workload pending in eachworker’s backlog, and how long they have been pending(as tracked by the system);2.
Worker performance : each worker’s productivity based onpast observation data;3.
System-level preference : the emphasis, given by the sys-tem operators, on achieving high collective productivitycompared to allowing for workers to be more rested; and4.
Personal preference : each worker’s mood at a giventime which is used to infer the level of productivitythat can be expected from the worker in the immedi-ate future following the happy-productive worker thesis(Oswald, Proto, and Sgroi 2015).Based on this index, CPL dynamically determines the tim-ing and amount of work a worker should perform in or-der to conserve the collective effort output while maintain-ing a high level of collective productivity in the system. Itcan be implemented in a distributed manner as a personalscheduling agent with a computational time complexity of O (1) . This enables it to effectively support the need forreal-time scheduling for large-scale workforce management.PL allows human values to be codified and algorithmicallyguide the recommendations by the scheduling agent to bal-ance worker wellbeing and system throughput. Through ex-tensive numerical experiments based on a large-scale real-world dataset containing over 5,000 workers’ performancecharacteristics, CPL is shown to significantly outperform al-ternative approaches, consistently achieving superlinear col-lective productivity (Sornette, Maillart, and Ghezzi 2014). Related Work
Existing AI-powered workforce management approaches(Lee et al. 2015) can be divided into two main categories:(1) the direct approaches and (2) the indirect approaches.Some direct approaches seek to balance the di-vision of labour among workers in a situation-aware manner through data-driven deliberation(Mason and Watts 2012; Yu et al. 2013a; Dai et al. 2013;Tran-Thanh et al. 2014b; Yu et al. 2015; Yu et al. 2016;Grossmann, Brienza, and Bobocel 2017; Yu et al. 2017a).Others design reputation and/or incentive mechanismsto motivate workers to work harder (Yu et al. 2011;Mao et al. 2013; Yu et al. 2013b; Tran-Thanh et al. 2014a;Miao et al. 2016; Zeng, Tang, and Wang 2017). They gen-erally leave it up to the workers to plan their rest breaks.There are approaches which implicitly limit how longa worker can continuously work by setting a budget onthe number of tasks they are allocated (Chen et al. 2015;Zenonos, Stein, and Jennings 2016). However, these ap-proaches do not opportunistically take advantage of periodswhen a worker a highly productive (e.g., periods of goodmood in our approach). Such ad hoc planning may nothelp the system maintain high collective productivity. WithCPL, a desirable balance between worker wellbeing andcollective productivity can be achieved.Indirect approaches (Morris, Dontcheva, and Gerber 2012)seek to induce good mood among workers in order toimprove their productivity. This is based on the happy-productive worker thesis (Oswald, Proto, and Sgroi 2015),which suggests that good (bad) mood improves (hampers)productivity. They also do not explicitly recommend restingto workers.A recent work (Yu et al. 2017b) is starting to explore howto opportunistically schedule rest. However, it does not ac-count for how long tasks have been pending in workers’backlog and can lead to schedules which do not make busi-ness sense. Without considering the urgency of the tasks, itis vulnerable to workers misreporting their mood. With theconceptual queue technique developed in this paper, CPLconsiders workers’ wellbeing, system objectives and situa-tional factors in a more holistic manner, and is better able todeal with misbehaving workers trying to game the system.
The Proposed
CPL
Scheduling Approach
The system model in this paper consists of a set of N work-ers and a set of M tasks at any given time slot t . • Workers are associated with personal profiles. Each pro-file contains information on a worker i ’s task backlogqueue q i ( t ) and his maximum productivity µ max i which indicates the maximum workload he can complete in agiven time slot t . The granularity of t can be adjusted tofit any given system’s requirements. The ground truth of µ max i may not be directly observable. With analytics toolssuch as Turkalytics (Heymann and Garcia-Molina 2011),it can be tracked and estimated statistically. • Tasks are associated with task profiles. The most impor-tant information for our purpose is the task deadline whichis expressed in terms of number of time slots since taskcreation before which a given task must be completed.The number of workers and tasks available in a given systemat different time slots may differ.We model the dynamics of a worker i ’s task back-log queue following previous research (Yu et al. 2013a;Yu et al. 2015): q i ( t + 1) = max[0 , q i ( t ) + λ i ( t ) − µ i ( t )] (1)where λ i ( t ) is the number of new tasks delegated toworker i during time slot t ; and µ i ( t ) is the num-ber of tasks completed by worker i during time slot t .Based on the ‘happy-productive worker’ thesis which sug-gests that productivity is positively correlated to mood(Oswald, Proto, and Sgroi 2015), µ i ( t ) can be expressed asa function of i ’s current mood and how much effort hespends on the tasks: µ i ( t ) = µ ( ξ i ( t ) , m i ( t )) = ⌊ ξ i ( t ) m i ( t ) µ max i ⌋ (2)where ξ i ( t ) ∈ [0 , is the normalized effort worker i spends during time slot t . m i ( t ) ∈ [0 , is worker i ’smood during time slot t , where 1 denotes the most positivemood. It can be self-reported by a worker using tools suchas MoodPanda ( http://moodpanda.com/ ), or throughfacial image analytics (e.g., in-vehicle monitoring of drivers’mood (Zimasa, Jamson, and Henson 2017)). Here, we treat m i ( t ) as an external parameter to CPL and do not assumethat the system is capable of predicting future mood values. Deriving the Work-Rest Index
In order to take task pending time into account, we propose a conceptual queue technique. Let Q i ( t ) denote a conceptualqueue which is being managed by a CPL agent on behalf ofa worker i . The queuing dynamics of this conceptual queueis defined as: Q i ( t +1) = max[0 , Q i ( t )+ µ max i [ q i ( t ) > & µ i ( t )=0] − µ i ( t )] (3)where [ condition ] is an indicator function. Its value is 1 if andonly if [condition] is satisfied; otherwise, it evaluates to 0.When first created, a conceptual queue is empty. The con-ceptual queue starts from 0 at t = 0 (i.e. Q i (0) = 0 ). Itssize increases by µ max i if and only if worker i ’s pending taskqueue is not empty at time slot t and worker i rests duringthis time slot. Its size decreases in the same way as q i ( t ) .For simplicity of notation, we denote µ max i [ q i ( t ) > & µ i ( t )=0] as x i ( t ) . Then, equation (3)can be re-expressed as: Q i ( t + 1) > Q i ( t ) + x i ( t ) − µ i ( t ) . (4)y re-arranging the above inequality, we have: Q i ( t + 1) − Q i ( t ) > x i ( t ) − µ i ( t ) . (5)Summing both sides of the above inequality over all t ∈{ , ..., T − } yields: T − X t =0 [ Q i ( t + 1) − Q i ( t )] > T − X t =0 [ x i ( t ) − µ i ( t )] . (6)Thus, we have: Q i ( T ) − Q i (0) > T − X t =0 [ x i ( t ) − µ i ( t )] . (7)Since Q i (0) = 0 , the above inequality is simplified as: Q i ( T ) T > T T − X t =0 x i ( t ) − T T − X t =0 µ i ( t ) . (8)From equation (8), it can be observed that the effect of theconceptual queue is to signal the scheduling approach as towhen the need to reduce pending workload shall take prece-dence over helping workers plan their rest. By ensuring thatthe computed µ i ( t ) values satisfy the queueing stability re-quirement of T P T − t =0 µ i ( t ) > T P T − t =0 x i ( t ) for the con-ceptual queue, a scheduling approach will ensure that tasksdo not stay pending indefinitely.By jointly considering q i ( t ) and Q i ( t ) , the Lyapunovfunction (Neely 2010) which measures the overall concen-tration of work among workers in a given system during timeslot t is: L ( t ) = 12 N X i =1 [ q i ( t ) + Q i ( t )] . (9)A large value of L ( t ) indicates that tasks are highly con-centrated among a small number of workers and/or thattasks have been pending for a long period of time. Bothof these scenarios are undesirable from a system produc-tivity perspective and shall be avoided as much as possi-ble. The constant term is included to simplify subsequentderivations without affecting the physical meaning of theformulation. We adopt the time-averaged Lyapunov drift , ∆ = T P T − t =0 [ L ( t + 1) − L ( t )] , to measure the changes inthe degree of seriousness of these two scenarios over time.We formulate a joint { effort output + drift } optimizationobjective function as: φ E { ˜ ξ ( t ) | ˜ q ( t ) , ˜ m ( t ) } + ∆ (10)which shall be minimized. φ > controls the empha-sis placed on conserving worker effort compared to gettingmore work done. A larger φ value can be interpreted asstronger emphasis on allowing workers to rest more. Thisvalue can be set by the system operators to express system-level preference on how to utilize the collective productivityof the workers. ˜ ξ ( t ) , ˜ q ( t ) and ˜ m ( t ) are vectors containingthe workers’ effort output values, the backlog queue sizes,and the self-reported mood during time slot t , respectively. Based on equation (1) and equation (9), the time-averagedLyapunov drift can be expressed as: (11) ∆ = 1 T T − X t =0 N X i =1 (cid:20)(cid:18) q i ( t + 1) − q i ( t ) (cid:19) + (cid:18) Q i ( t + 1) − Q i ( t ) (cid:19)(cid:21) = 1 T T − X t =0 N X i =1 (cid:20)(cid:18)
12 max[0 , q i ( t ) + λ i ( t ) − µ i ( t )] − q i ( t ) (cid:19) + (cid:18)
12 max[0 , Q i ( t )+ µ max i [ q i ( t ) > & µ i ( t )=0] − µ i ( t )] − Q i ( t ) (cid:19)(cid:21) T T − X t =0 N X i =1 (cid:20)(cid:18) q i ( t )[ λ i ( t ) − µ i ( t )] − µ i ( t ) λ i ( t )+ 12 [ λ i ( t ) − λ i ( t ) µ i ( t ) + µ i ( t )] (cid:19) + (cid:18) Q i ( t )[ µ max i [ q i ( t ) > & µ i ( t )=0] − µ i ( t )]+ 12 [( µ max i ) [ q i ( t ) > & µ i ( t )=0] − µ max i [ q i ( t ) > & µ i ( t )=0] µ i ( t )+ µ i ( t )] (cid:19)(cid:21) . This formulation enables simultaneous modelling of the ab-solute sizes of the real and the conceptual queues, the distri-bution of congestions among these queues, and the fluctua-tions of these queues over time. All three quantities shouldbe minimized in order to optimize our design objectives.In this way, we translate system-level efficiency and workerwellbeing requirements into queueing system stability con-cepts. Then, through maintaining queue system stability,CPL achieves these desirable objectives.Since neither λ i ( t ) nor µ i ( t ) can be infinite in real-worldsystems, we can simplify ∆ as: (12) ∆ T T − X t =0 N X i =1 (cid:20)(cid:18) q i ( t )[ λ i ( t ) − µ i ( t )] − µ i ( t ) λ i ( t )+ 12 [ λ + µ ] (cid:19) + (cid:18) Q i ( t )[ µ max [ q i ( t ) > & µ i ( t )=0] − µ i ( t )]+ 12 [ µ [ q i ( t ) > & µ i ( t )=0] + µ ] (cid:19)(cid:21) . where λ max > λ i ( t ) and µ max > µ i ( t ) for all i and t areconstant values. As we only aim to influence µ i ( t ) with rec-ommendations, we only retain terms containing µ i ( t ) . In thisway, by substituting equation (12) into equation (10), we ob-tain the following objective function:inimize: T T − X t =0 N X i =1 ξ i ( t )[ φ − ( q i ( t ) + Q i ( t )) m i ( t ) µ max i ] (13)Subject to: ξ i ( t ) , ∀ i, ∀ t (14) µ ( ξ i ( t ) , m i ( t )) µ max i , ∀ i, ∀ t (15)By minimizing equation (13) subject to Constraints (14) and(15), we simultaneously minimize the time-averaged totalworker effort output while maximizing the time-averagedcollective productivity in a given system. To minimize equa-tion (13), at each time slot t , we need to compute the val-ues of the expression [ φ − ( q i ( t ) + Q i ( t )) m i ( t ) µ max i ] forall i . For simplicity of notation, we denote [ φ − ( q i ( t ) + Q i ( t )) m i ( t ) µ max i ] as the Work-Rest Index (WRI) , Ψ i ( t ) .This index enables a scheduling agent to efficiently searchthrough a very large solution space to determine if the cur-rent situation is more suitable for work or rest. Opportunistic Work-Rest Scheduling
Algorithm 1 presents a distributed implementation of theCPL approach as a scheduling agent for each worker. Fora worker i , if Ψ i ( t ) < , it sets ξ i ( t ) = min h , q i ( t ) m i ( t ) µ max i i and computes the corresponding µ i ( t ) value; otherwise, itadvises the worker to rest for the current time slot. The com-putational time complexity of Algorithm 1 is O (1) , makingit highly scalable. The algorithm implements computation-ally the intuition that the more pending tasks in a worker’sbacklog queue, the longer these tasks have been pending,the more emphasis on worker wellbeing by system opera-tors, and the higher the worker’s current mood, the moreeffort should be expended towards completing tasks (subjectto the physical limitations of the worker’s effort output pertime slot) . Algorithm 1
CPL
Require: φ , q i ( t ) , µ max i and m i ( t ) . if Ψ i ( t ) < then ξ i ( t ) = min h , q i ( t ) m i ( t ) µ max i i ; Compute µ i ( t ) according to equation (2); else ξ i ( t ) = 0 ; µ i ( t ) = 0 ; end if return µ i ( t ) ;WRI incorporates the ethical considerations(Yu et al. 2018) by prioritizing worker wellbeing con-siderations through recommending opportunistic restbreaks, and allowing stakeholders to influence the AIrecommendations by expressing their preferences through φ and m i ( t ) . The recommendations from CPL are givento a worker in the form of the number of tasks he shouldcomplete over a given time slot (a 0 value indicates that theworker should rest) so as to make it actionable enough forthe worker to follow. Experimental Evaluation
To evaluate the performance of CPL under realistic set-tings, we compare it against four alternative approachesthrough extensive simulations. The characteristics of workeragents in the simulation are derived from the
Tianchi dataset( http://dx.doi.org/10.7303/syn7373599 ) re-leased by Alibaba. This real-world dataset contains infor-mation regarding 5,547 workers’ reputation (i.e. quality ofwork) and maximum productivity (i.e. the maximum num-ber of tasks a worker can complete per time slot). Thisdataset allows us to construct realistic simulations.
Experiment Settings
The five comparison approaches are:1. The
Max Effort (ME) approach: under this approach, aworker agent i always works as long as there are tasks inits backlog queue regardless of its mood.2. The Mood Threshold (MT) approach: under this approach,a worker agent i works whenever m i ( t ) > θ and thereare tasks in its backlog queue, where θ ∈ [0 , is a pre-determined mood threshold used by MT.3. The Mood and Workload threshold (MW) approach: thisapproach jointly considers a worker agent i ’s currentmood and workload to determine how much effort to ex-ert. Whenever q i ( t ) µ (1 , m i ( t )) > µ max i µ (1 , θ ) , worker i exerts up to the maximum effort subject to there beingenough tasks in its backlog queue, where θ ∈ [0 , is apredetermined mood threshold used by MW.4. The Affective Crowdsourcing (AC) approach(Yu et al. 2017b) which is similar to CPL but doesnot take task pending time into account.5. The
CPL approach proposed in this paper.The approach used to delegate tasks to worker agents un-der all comparison approaches is SMVM (Yu et al. 2017a).It dynamically distributes tasks among workers in asituation-aware manner in order to avoid over concentra-tion of workload. At each time slot, SMVM determines howmany tasks to delegate to each worker agent i in the simula-tions (i.e. SMVM computes λ i ( t ) for all i and t ) based on itscurrent reputation and workload. The principle implementedby SMVM is that the higher a worker agent’s reputation andthe lower its current workload, the more tasks should be del-egated to it. SMVM can be replaced by any other approach(Ho, Jabbari, and Vaughan 2013; Basu Roy et al. 2015) aslong as such an approach can determine the values of λ i ( t ) for all i and t .In order to create different experiment conditions, wevary the value of φ between 5 and 100 in increments of5. The values of θ and θ are varied between 0.05 and 1in increments of 0.05. The system workload is measuredin relation to the maximum collective productivity of theworker agent population, Ω = P Ni =1 r i µ max i . In this equa-tion, r i is a worker agent i ’s reputation and N = 5 , . Weadopt the concept of load factor (LF) from (Yu et al. 2016;Yu et al. 2017a) to denote the overall workload placed on thesystem. It is computed as the ratio between the number ofew tasks delegated to the worker agents during time slot t , W req ( t ) , and the maximum collective productivity Ω ofthe system (i.e. LF = W req ( t )Ω ). We vary LF between 5% to100% in 5% increments.Throughout the experiments, the mood for each workeragent i during time slot t , m i ( t ) , is randomly generated inthe range of [0 , following a uniform distribution. Thiseliminates the possibility for any of the comparison ap-proaches to predict a worker agent’s future mood basedon previous observations, thereby focusing the experimentalcomparisons on the effectiveness of the scheduling strate-gies. Under each LF setting, the simulation is run for T =10 , time slots. All tasks must be completed within 3 timeslots after they have been delegated. Evaluation Metrics
The performances of the five approaches in the experimentsare compared using the following metrics:1. The time-averaged worker effort output, ¯ ξ = T N P T − t =0 P Ni =1 ξ i ( t ) . The smaller the ¯ ξ value, thebetter the performance of an approach.2. The time-averaged task expiry rate, ¯ e = T N P T − t =0 P Ni =1 n ( e ) i ( t ) q i ( t ) , where n ( e ) i ( t ) is the num-ber of tasks in i ’s backlog which passed their deadlinesduring a given time slot t . The smaller the ¯ e value, thebetter the performance of an approach.3. The time-averaged task completion rate, ¯ µ = T P T − t =0 P Ni =1 µ i ( t ) N total ( t ) , where N total ( t ) is the total numberof tasks waiting to be completed in the system during agiven time slot t . The larger the ¯ µ value, the better theperformance of an approach.Since worker agents under the ME approach consistentlyexpend the most effort and achieve the highest task com-pletion rate, we use ME as the baseline for comparing theperformance of other approaches under different LF, φ , σ , θ and θ settings. Results and Discussions
Figures 1(a)–1(d) show the time-averaged task expiry ratesachieved by MT, MW, AC and CPL respectively under vari-ous experiment settings. As MT uses mood as the thresholdto control worker effort output, the changes in task expiryare directly related to mood values (Figure 1(a)). On aver-age, 29% of the tasks under MT expire before they can becompleted. MW is also a threshold-based approach. How-ever, its threshold consists of a combination of workers’mood and their current workload. Thus, its task expiry rateincreases with both mood and LF with the effect of moodbeing more significant (Figure 1(b)). On average, 30% ofthe tasks under MW expire before they can be completed.AC is not a threshold-based scheduling approach. A workercan indicate to AC his desire to rest by adjusting the valueof the control variable σ . As σ and LF values increase, anincreasing percentage of tasks expire under AC. On average,7.5% of the tasks under AC expire before they can be com-pleted, which is significantly lower than MT and MW. As AC only considers mood and workload when making work-rest recommendations, workers may fall into the conditionin which their mood and workload trigger AC to recommendresting. However, their workload is also not low enough tocause the task delegation approach to delegate new tasks tothem. Therefore, AC continues to recommend resting un-til pending tasks pass their deadlines and become expired.This problem is addressed by CPL as it takes task pendingtime into account with the conceptual queue technique whenoptimizing work-rest scheduling. As shown in Figure 1(d),increases in φ and LF values result in an increasing percent-age of tasks expire under CPL. The task expiry rate underCPL is lower than that under AC. On average, 5.4% of thetasks under CPL expire before they can be completed, whichis significantly lower than MT, MW and AC considering thescale of the experiment.The time-averaged effort output ¯ ξ achieved by all five ap-proaches is shown in Figure 1(e). Compared to ME, all otherapproaches achieved significant savings in effort as LF in-creases. This is partially due to the SMVM task delegationapproach used in the simulations. When LF is low, tasks aremostly concentrated on worker agents with good reputation.In this case, the task backlogs of individual worker agentswho have been delegated tasks tend to be relatively high,which makes scheduling approaches allocate less time forthese workers to rest in order to meet task deadlines. AsLF increases, the workload is spread more evenly amonga larger segment of the worker agent population, creatingmore opportunities for scheduling approaches to slot in restbreaks. The ¯ ξ values achieved by MT, MW and AC sta-bilize between 20% and 40% while that of CPL stabilizesaround 60%. MW achieves the lowest worker agent effortoutput. The time-averaged task completion rates ¯ µ achievedby all five approaches are shown in Figure 1(f). It can beobserved that CPL achieves the highest ¯ µ values which sta-bilize around 85% and are higher than AC, MT and MW by22%, 90% and 108%, respectively.Figure 1(g) shows the performance landscape of MT,MW, AC and CPL as a percentage of ME under differentcontrol parameter settings. MW and MT both use moodthresholds ( θ and θ , respectively) to control effort output.The higher the mood threshold values, the lower the effortoutput (and hence the task completion rates) achieved byMW and MT. On the other hand, mood serves as one of theinputs to AC and CPL. Both AC and CPL allow a workeragent to specify a general emphasis on conserving effort us-ing variables σ and φ , respectively. The larger the values ofthese variables, the more emphasis is placed on effort con-servation. By varying the φ value in CPL, we can control thetrade-off between worker effort output and task completionrate (from spending 92% of the ME effort output and achiev-ing 99% of the ME task completion rate, to spending 53%of the ME effort output and achieving 78% of the ME taskcompletion rate). CPL consistently and significantly outper-forms both MT and MW, conserving significant worker ef-fort while achieving high task throughput. In effect, CPLlimits the range of trade-off between work and rest achievedby AC based on consideration of an additional situationalfactor – the task pending time – in order to achieve better θ A v g . T a sk E x p i r y R a t e ( % ) θ A v g . T a sk E x p i r y R a t e ( % ) σ A v g . T a sk E x p i r y R a t e ( % ) φ A v g . T a sk E x p i r y R a t e ( % ) A v g . E ff o r t O u t pu t ( % ) MEMWMTACCPL A v g . T a sk C o m p l e t i on R a t e ( % ) MEMWMTACCPL σ =50 σ =100 θ =0.3 θ =0.5 θ =0.3 θ =0.5 φ =50 φ =100% of ME Effort Output(g) % o f M E T a sk C o m p l e t i on CPLACMTMW % o f M E T a sk C o m p l e t i on Figure 1: Experiment results: (a) the time-averaged task expiry rates achieved by MT under various θ and LF settings; (b) the time-averaged task expiry rates achieved by MW under various θ and LF settings; (c) the time-averaged task expiry ratesachieved by AC under various σ and LF settings; (d) the time-averaged task expiry rates achieved by CPL under various φ andLF settings; (e) comparison of the time-averaged effort output achieved by various approaches under different LF settings; (f) comparison of the time-averaged task completion rates achieved by various approaches under different LF settings; (g) the time-averaged task completion rates vs. the time-averaged effort output achieved by various approaches under different parametersettings as a percentage of those achieved by ME; (h) the time-averaged task completion rate vs. the time-averaged effort outputachieved by various approaches averaged over different parameter settings as a percentage of those achieved by ME.collective performance. In the worst case scenario in which θ and θ are set to 1, indicating that workers are unwillingto work under any mood condition, the worker effort outputand the task completion rates achieved by both MT and MWare 0. This is expected as in MT and MW, mood is used asthe threshold to control effort output. However, under AC,mood is only one of the situational factors considered by theapproach. CPL adds in task pending time on top of the sit-uational factors used by AC to make scheduling decisions.Even under scenarios in which φ is set to 100 (indicating thatworkers place very high emphasis on rest), CPL is still ableto maintain a long-term average effort output of 53% takingadvantage of favourable working conditions whenever pos-sible to achieve average task completion rates of about 78%.This is more advantageous from a business perspective com-pared to the worst case performance by AC of about 50%.When we compute the averages of the results shown inFigure 1(g) over their respective setting variables (i.e. φ , σ , θ and θ ), we obtain Figure 1(h) showing the overview oftheir performances. The diagonal dotted line represents lin-ear productivity, meaning that an increase in effort outputresults in a directly proportional increase in collective pro-ductivity. It can be observed that ME, MT and MW all fallon the linear productivity line, whereas AC and CPL are sig-nificantly above this line in a region of superlinear produc-tivity (Sornette, Maillart, and Ghezzi 2014). Under AC andCPL, an increase in effort output results in a disproportion-ally larger increase in collective productivity, indicating thatthe collective productivity achieved is larger than the sum of individual workers’ productivity. Overall, CPL achieves89% of ME task completion rate with 69% of the ME workereffort output, which is the most desirable work-rest trade-offamong the five approaches from a system perspective. Conclusions and Future Work
Improving collective productivity is an important problemfacing many social and economic systems. How to dynam-ically adapt workers’ work-rest schedules in response tochanging situations in order to maintain a high level of pro-ductivity and worker wellbeing remains an open researchquestion. The proposed CPL approach translates consider-ations on workers’ mood, workload and pending time of thetasks in their backlogs into actionable personalized work-rest schedules. It establishes a framework to model com-plex relationships between work and rest, and helps work-ers optimize the balance between work and rest in order toachieve superlinear collective productivity. Taking into ac-count its polynomial time complexity, CPL is an effectiveand scalable approach to help workers benefit from oppor-tunistic rest. By nudging workers to be ‘lazy’ at opportunetimes, CPL achieves collective productivity which is largerthan the sum of individual productivity. It provides a wayto design ethically aligned workforce management systemsthat sustain long-term effective participation by promotingproductive laziness among workers.In future research, we plan to testbed CPL in a crowd-sourcing platform (Pan et al. 2016) to reach out to more di-erse users and study how to improve the approach in thepresence of various behaviour patterns and how to fostertrust (Shen et al. 2011) with user by explaining the rationalebehind the recommendations.
Acknowledgments
This research is supported, in part, by Nanyang Technological Uni-versity, Nanyang Assistant Professorship (NAP) and the NationalResearch Foundation, Prime Minister’s Office, Singapore under itsIDM Futures Funding Initiative.
References [Basu Roy et al. 2015] Basu Roy, S.; Lykourentzou, I.; Thirumuru-ganathan, S.; Amer-Yahia, S.; and Das, G. 2015. Task assignmentoptimization in knowledge-intensive crowdsourcing.
VLDB Jour-nal
International Journal of Crowd Science
IJCAI , 1113–1119.[Dai et al. 2013] Dai, P.; Lin, C. H.; Mausam; and Weld, D. S. 2013.POMDP-based control of workflows for crowdsourcing.
ArtificialIntelligence
CSCW ,628–638.[Grossmann, Brienza, and Bobocel 2017] Grossmann, I.; Brienza,J. P.; and Bobocel, D. R. 2017. Wise deliberation sustains coopera-tion.
Nature Human Behaviour
WWW , 477–486.[Ho, Jabbari, and Vaughan 2013] Ho, C.-J.; Jabbari, S.; andVaughan, J. W. 2013. Adaptive task assignment for crowdsourcedclassification. In
ICML , I–534–I–542.[Lee et al. 2015] Lee, M. K.; Kusbit, D.; Metsky, E.; and Dabbish,L. 2015. Working with machines: The impact of algorithmic anddata-driven management on human workers. In
CHI , 1603–1612.[Leiter and Maslach 2015] Leiter, M. P., and Maslach, C. 2015.Conquering burnout.
Scientific American Mind
HCOMP , 94–102.[Mason and Watts 2012] Mason, W., and Watts, D. J. 2012. Collab-orative learning in networks.
Proceedings of the National Academyof Sciences USA
Decision Support Systems
Science
IEEE Internet Computing
Stochastic Network Optimizationwith Application to Communication and Queueing Systems . Mor-gan and Claypool Publishers.[Oswald, Proto, and Sgroi 2015] Oswald, A.; Proto, E.; and Sgroi,D. 2015. Happiness and productivity.
Journal of Labor Economics
AAAI , 4248–4249.[Shen et al. 2011] Shen, Z.; Yu, H.; Miao, C.; and Weng, J. 2011.Trust-based web-service selection in virtual communities.
Web In-telligence and Agent Systems ⊞ . : Superlinear productivity in collectivegroup actions. PLoS ONE
AAMAS , 477–484.[Tran-Thanh et al. 2014b] Tran-Thanh, L.; Stein, S.; Rogers, A.;and Jennings, N. R. 2014b. Efficient crowdsourcing of unknownexperts using bounded multi-armed bandits.
Artificial Intelligence
Proceedings of the National Academy of Sciences USA
KDD , 905–913.[Yu et al. 2011] Yu, H.; Shen, Z.; Miao, C.; and Tan, A.-H. 2011. Asimple curious agent to help people be curious. In
AAMAS , 1159–1160.[Yu et al. 2013a] Yu, H.; Miao, C.; An, B.; Leung, C.; and Lesser,V. R. 2013a. A reputation management model for resource con-strained trustee agents. In
IJCAI , 418–424.[Yu et al. 2013b] Yu, H.; Shen, Z.; Leung, C.; Miao, C.; and Lesser,V. R. 2013b. A survey of multi-agent trust management systems.
IEEE Access
AAAI , 1305–1311.[Yu et al. 2016] Yu, H.; Miao, C.; Leung, C.; Chen, Y.; Fau-vel, S.; Lesser, V. R.; and Yang, Q. 2016. Mitigating herd-ing in hierarchical crowdsourcing networks.
Scientific Reports
Scientific Reports
IEEE ICA , 121–126.Yu et al. 2018] Yu, H.; Shen, Z.; Miao, C.; Leung, C.; Lesser, V. R.;and Yang, Q. 2018. Building ethics into artificial intelligence. In
IJCAI , 5527–5533.[Zeng, Tang, and Wang 2017] Zeng, Z.; Tang, J.; and Wang, T.2017. Motivation mechanism of gamification in crowdsourcingprojects.
International Journal of Crowd Science
AAAI , 3936–3942.[Zimasa, Jamson, and Henson 2017] Zimasa, T.; Jamson, S.; andHenson, B. 2017. Are happy drivers safer drivers? Evidence fromhazard response times and eye tracking data.