Conclusions and Suggestions for Future
Research
In this thesis dynamic inbound contact centers with heterogeneous agents and retrials of impatient customers were analysed. The term dynamic characterises the processes of inbound contact centers in two ways. On the one hand the processes of a real-world contact center are time-dependent and on the other they are random. The interarrival times, service durations and abandonment behaviour depend on the time of the day and the day of the week. Furthermore, customers contact the center at random times and the service durations and waiting time limits are random as well.
In order to model and analyse both aspects of dynamics, we used the so-called strong approximations which contain a fluid approximation and a dif-fusion refinement. By means of the fluid approach we showed how retrials and time-dependent parameters influence technical and economical performance measures. Furthermore, we used a diffusion refinement of the fluid approach to investigate the impact of the parameters on the variability of the queueing processes.
Finally, we developed an algorithm for integratively solving a staff require-ment planning and shift scheduling problem which emerges in contact centers with time-dependent processes and retrials. This algorithm relies on the fluid approximation.
Besides heterogeneous and impatient customers, differently skilled agent groups, time dependencies and randomness, in many real-world contact cen-ters the arrivals often consist of primary attempts and retrials of impatient customers. These retrials influence the performance of the contact center and can lead to under- or overestimation of the demand. In order to investigate the various impacts on the performance we use models with different kinds of customer requests and differently skilled agents. The customers are assumed to be impatient and a certain percentage of customers retries after having abandoned. Furthermore, the arrival rates strongly depend on time of the day.
In the traditional stationary queueing approach mainly used in practise just one aspect of dynamics can be considered. In this approach mainly the randomness of the arrivals, service completions and abandonments is ad-dressed. This aspect has an enormous influence in very small contact cen-ters, e.g., for emergency calls, while in big contact centers analysed in this thesis its influence diminishes. In the traditional approach the modelling of heterogeneous customers and agents as well as retrials is another serious prob-lem because the stochastic processes become processes of higher dimensions. Consequently, the generator becomes a matrix of matrices and the stationary distibution becomes nearly uncomputable as argued in Section3.1.4.
Therefore, we use the fluid approach. By means of this approach the in-fluence of the time-dependencies can be shown and is stressed. Furthermore, the distribution of the random events becomes negligible, i.e., it does not matter whether the service times are exponentially, lognormally or normally distributed. We show, that the retrials can easily be modelled and even dif-ferent customer classes and various groups of agents can be imbedded in this approach because the fluid approach leads to a simple first order und numeri-cally stable initial value problem. This initial value problem can be solved by simple methods like the Euler method. Each customer class leads to two dif-ferential equations of the initial value problem: one to determine the number of customers in the system and one to determine the number of customers in the orbit waiting to recall. Additional agent groups lead to additional terms in the differential equations which describe the supplemental non-preemptive priority rule. The modelling by the fluid approach gives rise to a preemptive priority rule. Numerical examples show that the error due to the change is very small. Unfortunately, not as many performance measures are calculable as for the stationary approach, e.g., the well-known X/Y service level can-not be calculated. For this performance measure we need the distribution of the waiting time. However, we showed that only few performance measures are needed to analyse and benchmark the performance of a contact center. We derive the average waiting time, the probability of being served and the utilisation as technical performance measures. These measures are used to calculate the profit as an economical performance measure.
We investigate the influence of the service rate, the abandonment rate, number of agents, the mean time to retrial and the probability of retrial on the performance measures. Thereby we focus on the influence of the retrial parameter. Additionally, we compare the results to simulation results to show that the approximation is very accurate. Higher service rates and a greater number of agents lead to smaller waiting times, a higher probability of be-ing served and less utilisation. The abandonment rates solely influence the waiting time such that higher rates leads to smaller waiting times. We show that the probability of being served and the utilisation are not effected. The influence of the mean time to retrial is shown to be smaller than the influence of the probability of retrial. The shorter mean times to retrial lead to a little
longer waiting times in the case of high load. The influence on the probability of beeing served is similar. Higher probabilities of retrial give rise to higher waiting times and smaller probabilites of being served. Furthermore, we show that the retrial parameter do not influence the utilisation because retrial are solely caused, if the contact center is temporarily overloaded.
The diffusion refinement is based on the fluid approach. It considers the randomness of the processes. All random processes are approximated by nor-mally distributed processes with a mean given by the fluid approach. That is why, some shortcomings of the fluid approach especially for the times of critical loading apply to the diffusion approximation as well. The diffusion refinement is needed to extend the initial value problem of the fluid approach for the variances and covariances of the processes describing the number of customers.
These differential equation for the variances and covariances are difficult to derive. In order to solve the extended initial value problem we had to use a Runge-Kutta-Method because the differential equations of the variances and covariances are quite sensitiv to the service and abandonment rates. If the ser-vice rate and the abandonment rate differ a lot, the variances and covariances change drastically. It was shown that the calculated variances and covari-ances are useful for the analysis of interdependencies of different customers processes. They might be of value to get more robust schedules of agents in a future optimisation approach. However, for a staff requirement planning and shift scheduling approach we do not need the diffusion refinement. This can be done solely by means of the fluid approximation.
We have modelled contact centers with retrials and time-dependent pro-cesses by means of fluid approximation and its diffusion refinement. Thereby we included the heterogeneity of customer requests as well as the different groups of agents. We restrict the model to the case of two customer classes and three agent groups. However, we showed that this model can easily be derived from the simple model and more complex models can be deduced as easily. We show that the number of customers in the system and the orbit are positive correlated. The correlation reduces if the contact center becomes more and more overloaded, i.e., the processes of the number of customers in the system and the number of customers waiting to retry become more and more independent.
The simulation we used is based on the traditional contact center model which assumes exponentially distributed interarrival and service times as well as waiting time limits and times to retrial. The advantages and disadvantages of this assumptions were discussed in Section2.2. Therefore, other distribu-tions for the service times, waiting time limits, and times to retrial should be simulated and the results should be compared to the approximation. Further-more, in the simulation each priority rule can easily be simulated while the fluid approximations leads to a preemptive priority rule. We showed that the
error due to this change in the priority rule is negligible small.
In addition the results of the approximation should be compared to the data of a real-world contact center. In order to collect and evaluate statis-tical significant retrial data it is necessary to mark a big number of calling and abandoning customer such that one can identify recalls. However, many customers may call several times from different telephones or locations. There-fore, it is difficult to identify a recall such that it is hard to get good data.
For the short-term economic optimisation and operational planning, the allocation of shifts and the agents associated with these shifts is a major prob-lem. We have shown how the fluid approach can be incorporated in an optimi-sation problem to solve the staff requirement planning and a shift scheduling problem simultaneously. For this purpose, we have derived a general profit function for the contact center model which is maximised.
We showed that the structure of the profit function is quite regular and concave almost everywhere. Only few regions of the profit function seem to be non-concave. However, this discrepency might be due to numerical instabilities in the computation.
We used an initial procedure and simulated annealling to solve the op-timisation problem. This simulated annealling heuristic is very simple and questionable because it does not converge to the optimal solution for sure. The parameters of the simulated annealling algorithm must be chosen very carefully and depend on the problem itself. Therefore, the implementation of other algorithms should be a fruitful task for future research.
The initial procedure is based on the relative profit margin of an addi-tional shift. The profit margins are calculated from the arrival rates and the service rates of the different agents weighted by the revenues and costs. We have shown that this procedure of the staffing algorithm already leads to a very accurate solution, i.e. high performance and profit. As the simulated an-nealling heuristic starts with this solution we get almost optimal results.
We have shown that the optimisation problem can be easily extended. Even more general problems with various customer classes and more agent groups as discussed in Section 5.5 can be solved. A further customer class leads to at least two additional differential equations in the fluid approach and additional terms in the profit function which all have the same structure. One differential equation is needed to describe the number of customers in the system and one for the number of customers in the orbit. All differential equations have the same structure as well. With the fluid approach we are also able to model customers, who change their customer class. An example of such a change is a customer who first tries to phone and afterwards writes an e-mail.
If another agent group is added, we have to decide which customer class is served first and which second. Furthermore, the order by which customers are
referred to the agent groups has to be determined. This leads to another term or a change in the terms of the differential equations. The number of differen-tial equations remains constant and the new terms have a similar structure as the terms, which describe the service by agents and the priority of customers. However, it will be necessary to improve the scheduling algorithm. Other optimisation algorithms may lead to even better results or find them more quickly. Furthermore, we used a fixed set of predefined shifts with fixed rest breaks where in practise more flexibility may be needed. Therefore, one should include more flexible rest breaks and shifts, e.g., one could consider time win-dows for rest breaks. In our model we assumed that agents are available to work according to each scheduled shift. However, in real-world contact cen-ters the contracts of employment and the preferrences of the agents lead to additional constraints. Therefore, it would be worthwhile to enlarge the shift scheduling to a tour scheduling problem which includes day-off scheduling.
Finally, another extension could involve the results of the diffusion refine-ment in order to make the staffing and scheduling decision more robust with respect to the variability of the processes. However, this seems to be not as worthwhile in practise, because the diffusion approximation is difficult to de-rive and the benefits are small in a big contact center. The reason is that the effects of randomness become more and more negligible in big contact centers. Therefore the major focus should lie on improving the staffing and scheduling algorithm. By means of the methods presented in this thesis and improved staffing and shift scheduling algorithms the calculation of shift scheduling be-comes much quicker and easier. That is why an integration of the methods into a software tool will be worthwhile for the management of call and contact centers.
