Figure 5.18:Total time of containers that are delivered outside time window with different time windows
Figure 5.18 shows the results for different stop criteria. We see that, compared to the normal situation, only a stop temperature yields worse performance for the ‘Total time outside TW’. When the stop criterion is the number of temperature levels without improvement (Exp. 20), we have a better performance in the ‘No decoupling’ scenario, but not in the other two scenarios.
5.6 Conclusions of solution tests
This section answered the research question ‘What are the best settings to improve the perfor- mance with using decision support for container transport scheduling?’ Section 4.3 described solution model we test this model in this chapter.
• Section 5.1 described the experimental set-up that we use to test the solution methods. This involves the assumptions, experimental factors and the planning scenarios. We test different optimization goals, initial solutions, improvement operators, and stop criteria. Each experiment is tested on four planning scenarios, which disable or force decoupling at customers, where there are no loading/discharge times, and the original situation. The data is gathered from CTT.
• Section 5.2 described the experiments. Not all combinations of experimental factors and planning scenarios are tested, because this is too time consuming.
• Section 5.3 verified the model and the settings of start an stop temperature and the cooling factor, by using the acceptance ratio and the performance curve.
• Section 5.4 showed the results of the experiments. The benchmark situations perform well for the loading/discharge time related KPIs. The difference between timeliness and costs related KPIs is clear in the time window related KPIs, in which the number of missed time windows decreases from about 55 to around 5. For the minimization of the number of trucks, the benchmark situation performs better than minimization of trucks optimization goal. The best results is 36 trucks in the benchmark case, and 37 when minimizing the trucks. The worst performance is when we optimize the total number of containers before the LD time, because there are 48 to deliver the containers in time. Minimizing the total travel time can be done by minimizing the travel time or minimizing the total time of
detours, which results in about 6.5 and 7 days of travelling, compared to almost 8.5 days in the best benchmark case. For the waiting time we can also optimize ‘Time after LD time’ or the time window related KPIs, but the minimization of waiting time still performs the best as expected. The total waiting time can be reduced from 3 days in the best benchmark case to about 16 hours in the normal situation. Reducing the number of detours and time of detours can be done best by reducing the time of detours. This reduces the number of detours from 129 to 78, and the time of detours from almost four days to almost 1.5 days. The results of these experiments should be further researched, because there are different temperatures used for the experiments, which means that there are different acceptance ratios.
• Section 5.5 test different settings of parameters. We see that only crossing of routes performs better than only moving or swapping jobs. A sequence of one job all the time performs better than changing from 3 to 2 to 1 using the Or-opt operator. Nevertheless, a combination of crossing, moving, and swapping performs the best in almost all situations. It is possible to deliver all jobs within a time window of 20 minutes before and after the LD time, but not within a 10 minutes time window. We see that, compared to the normal situation, only a stop temperature yields worse performance for the ‘Total time outside TW’ compared to the where both stop criteria should be true.
Chapter 6
Conclusions and discussion
"All models are wrong, but some are useful"
Box and Draper (1987)
The quote above illustrates the difficulties when building a model, which is a simplified real- isation of the complex reality. In the same way, assumptions are used in the solution method and thus influence the results. This chapter discusses the results and the relevance of the results compared with the complex reality of truck scheduling. We start the chapter with the conclu- sions of this research in Section 6.1. Section 6.2 describes the limitations of the model. Section 6.3 addresses recommendations for CTT and for related problems. This chapter ends in Section 6.4 with suggestions for further research.
6.1 Conclusions
This chapter summarizes the answers to the research question and the sub questions. The research goal of this thesis is to give support to the truck planners for container transport to improve the performance. Four research questions helped us to achieve this goal.
Chapter 2 described the context analysis. The container (transport) sector is growing from year to year, and this also influences the transport of containers in the hinterland. CTT faces this growth and the main problem is that it is difficult for the truck planners to use all this information of containers and bookings in their advantage. In an ideal situation, this information leads to a synchromodal network, in which the synchronization of information leads to better use of the modalities barge, train, and truck. The planning and scheduling for the modalities is done separately at CTT, but they need information of each other. Decisions regarding the allocation of containers to the barge or train influences the truck scheduling, but it is not clear for the barge and train planners what the consequences are of this decision. This may result in additional travel time and also that other containers are too late at the customer or sea terminal. Decision support can help to give insight in the network, such that these situations are made insightful.
We reviewed the literature about container transport in Chapter 3, and in particular focused on truck scheduling. The literature described the problem as a vehicle routing problem, but there is not a single type of vehicle routing problem which fully describes the complex situation at CTT. A classification scheme helped to describe the problem and we concluded that the problem at CTT has the following characteristics of a VRP:
• Vehicles with fixed capacities
• Heterogeneous fleet
• Time windows at customers, terminals and for resources
• Pick up and delivery
• Release and due dates
• Multi-depot
• Full truck load
The chapter continued to find a solution method for this problem. We searched for a heuristic, because heuristics are able to find good solutions in an acceptable amount of time. Simulated annealing turned out to be most applicable, because it is able to escape from local optima and does not have to calculate the performances of all its neighbors like in tabu-search. The heuristic comprises a construction phase and an improvement phase. In the first phase, the output is an initial truck schedule for one day. In the second phase, jobs are moved and swapped between trucks, aiming to find a schedule better than the initial schedule. The characteristic of simulated annealing is that it sometimes also accepts worse schedules than the current schedule, with a higher purpose to find a better solution at the end. Chapter 3 concluded with literature about implementing decision support. When our model is ready, validated, and verified, this does not necessarily mean that it will be accepted by the employees. The decision support system must at least fulfill three requirements, it should be: realistic and efficient, inexpensive to acquire and to modify, and easy to use and enhance the job.
Chapter 4 described the solution method. The chapter started with some use cases at CTT, which should be avoided in the future by using our decisions support methods. The most important and easiest, is the merging of a delivery and a pick up job, at locations nearby. This combination replaces two empty trips, one to each of the customers, by one trip from the delivery customer to the pick up customer.
The chapter continued with the solution methods for the different functionalities of a future decisions support system. The following functionalities are described:
• Scheduling based on departure time
• Calculate number of trucks needed
• Offline scheduling
• Online scheduling
• Synchromodal scheduling
Chapter 5 test the solution methods of Chapter 4. The focus in the experiments is on offline scheduling, which produces an initial schedule for one day. The key performance indicators are customer related (focus on timeliness) or related to CTT (total travel time, number of trucks, detours, and waiting time). The experiments focus on four different scenarios. The first scenario