Over the last 20 years, researchers have begun to apply simulation to operational problems in real-time, rather than system design problems. For example, Cheng (2007) considers the relocation of Fire Service vehicles when areas are left uncovered due to vehicles attending a major incident nearby. The vehicles need to be moved to reduce the expected fatality rates across the region. This is a combinatorially hard problem with legal restrictions that a solution must be found within one minute.
Some examples were discussed in Section 2.1.2 for airline disruption management. In the aviation industry, Scala et al. (2018) propose the use of simulation for scheduling aircraft movements at airport runways based on a rolling time window. One prevalent methodology for real-time decision making with simulation is symbiotic simulation.
This section introduces symbiotic simulation, briefly describes some applications and then discusses the issue of validating these systems.
2.4.1 Symbiotic Simulation
Symbiotic simulation is a paradigm involving a close interaction between a simulation model and the physical system it is modelling (Fujimoto et al., 2002; Aydt et al., 2008a). The simulation receives measurements and observations from the physical system to improve its representation. In turn, the physical system can use the infor-mation gained from the simulation analysis to improve its performance. A diagram of the interaction between the physical and simulation systems is shown in Figure 2.4.1.
When the simulation is triggered, it can evaluate a number of possible options through
Physical
System Observations Implementation
Control Output
Analysis
Simulation Simulation
Simulation What-If Experiments
Figure 2.4.1: A diagram of a symbiotic simulation system. Adapted from Fujimoto et al. (2002).
a simulation optimisation process to aid decision makers. In a fully automated system, the findings could be implemented automatically. The simulation could be used reac-tively, periodically (pro-actively) or to prevent a forecasted issue (Aydt et al., 2008b).
Aydt et al. (2009) review the important aspects of symbiotic simulation. Onggo et al.
(2018) describe a symbiotic simulation system as a hybrid system, linking simulation with many other fields such as Machine Learning, Optimisation, Statistics and areas of computer science such as data acquisition.
The work in this thesis is based on reactive triggering to disruptive events, though it may also have value in periodic use. Furthermore, the aim is to aid decision makers, and thus have a direct impact on the physical system, known as a ‘closed-loop’ system (Aydt et al., 2009). There are other forms of symbiotic simulation, as discussed by Aydt et al. (2009), but these are not discussed here.
There have been many and varied applications of symbiotic simulation within the
literature. The remainder of this section describes some of these, focussing on those acting as a decision support system.
Several papers consider the manufacturing sector. Aydt et al. (2011) apply sym-biotic simulation to a semiconductor manufacturing plant consisting of a series of stations, each with a set of machines. Periodically, the simulation is used to help schedule configuration changes at each station for the next period, depending on the demand forecast. Fanchao et al. (2009) apply a symbiotic simulation to a lubricant supply chain. Two problems are faced. The customer order problem is triggered when certain critical states are entered. The inventory control management is trig-gered periodically, aiming to find the minimal number of reorder points whilst meeting the demand in a dynamic way. Chiroma et al. (2018) describe the development and implementation of a symbiotic simulation system at an automotive manufacturing plant.
There are many applications of symbiotic simulation to the transport and logistics sector (Fujimoto et al., 2016). Vu et al. (2013) propose a symbiotic traffic simulation to aid traffic management in the event of a road incident. If an incident occurs, a set of potential paths for a car is simulated using traffic forecast information. However, if every car in the city used such a device and received the same advice, the congestion may simply be moved to a different part of the city. Aydt et al. (2012) propose the use of a similar system in which data from a group of cars could be used to globally optimise the traffic system by symbiotic simulation. Sunderrajan et al. (2016) describe a symbiotic simulation that uses GPS data from traffic to control a ramp-metering mechanism for managing the flow of traffic onto a junction. The simulation is a
macroscopic traffic model. Huang and Verbraeck (2009) consider how best to use new data from the physical system to perform an online calibration of a data-driven rail transport simulation system.
In a theoretical study one cannot test a symbiotic simulation system on a real physical system to evaluate its performance. Thus, the symbiotic simulations are generally run alongside another simulation designed to emulate the real world, acting as a proxy. In some cases, the proxy real world is more detailed and complex than the simulation used to make the decisions. For example, both Sunderrajan et al.
(2016) and Vu et al. (2013) use a microscopic traffic simulator as the proxy real world to their simulation, which are macroscopic and mesoscopic, respectively. Using a different simulation model mimics the practical implementation, where a simulation tool will not account for all aspects of the physical system. This concept is used in Chapter 7 when considering the evaluation of simulation-based decision support systems.
2.4.2 Validation of Real-time Simulation Systems
The validation of reusable simulation models is difficult, as the initial conditions usually play a very important role in how the simulation may behave. Oakley et al.
(2020) develop a symbiotic simulation to model bed occupancies across a hospital.
This is run periodically to predict occupancies over the next week. To validate the model, the authors propose a method that looks at the changes in bed occupancy over time, rather than the absolute values. Looking at the changes in the system means that one does not need to validate each possible initial condition against the limited
data that starts in the same state. The distribution of the change in occupancy seemed to be relatively stable across various initial conditions, making this an appropriate method for validation. The emphasis is focussed on reinforcing trust in the model rather than hypothesis testing.
If a reusable model is used over a long period of time, the model may need re-validation or re-calibration to reflect possible changes in the physical system (Aydt et al., 2011). Part of the power of symbiotic simulation is the ability to adapt the model based on new data from the real world (Onggo et al., 2018). A number of studies have looked into how to calibrate simulations to help accommodate this evolution.
Huang and Verbraeck (2009) and Huang et al. (2010) consider model validation and calibration of rail transit simulation models. The primary mechanism is to periodically compare the simulation and real world outputs and, if they deviate sufficiently, update the directly observable parameters (such as train location) with the real values, and use these to infer updates for the latent values (such as velocity and acceleration).
The authors suggest some mechanisms for this, but acknowledge that for stochastic output, this is much more difficult. This assumes that the model goes through a phase of validation before it is used to make predictions and decisions. A new validation phase commences if the state of the simulation deviates too much from the real world.
Papathanasopoulou et al. (2016) consider the online calibration of microscopic traf-fic simulators. The method proposed optimises the parameters of a (deterministic)
‘car-following’ model to minimise the difference between the observed and predicted speed. Unlike the static calibration method which would fit the parameters across large amounts of historical data, the online version re-fits the parameters at each time
step. The results show that this method allows greater prediction power for vehicle speed than the static approach. Hashemi et al. (2017) consider a similar system, but split the problem of updating parameters between multiple ‘agents’, who control a subset of the parameters. Rather than updating all parameters each time, a rein-forcement learning algorithm is used to let the agents decide whether or not to update the parameters, based on the current extent of the discrepancy and past rewards for the correction. This helps to prevent unnecessary model adjustments, reducing the number of false positives. In both cases, the experiments were on historical data rather than a real-world emulator.