SIM framework uses genetic algorithms to simulate typical weekday traffic. It starts with an initial demand, computed from several data sources, e.g. census data, questionnaires. This initial demand consists of complete activity chains for all agents for the whole day. This demand is then executed with a mobility simulation, summing up the experienced travel times delays and times of activity for every agent. After each iteration, new plans are being calculated, based on the results of the simulation run. Therefore a fitness function, which determines how the experienced travel/ac- tivity times should be rated, is defined and genetic algo- rithms generate new mutations of the executed plans with respect to this fitness function. The new plans will be exe- cuted in the simulation again. This process converges to a Nash equilibrium. The range of iterations necessary for the system to move towards a Nash equilibrium can vary from a few iterations up to several hundreds. Therefore reducing the execution time of the mobility simulation is of great in- terest. There are different implementations of the mobility simulation in the MATSIM framework. The most advanced is the Java-based implementation of the queue simulation algorithm . Although it has been tried to implement a multi-core version of the queue simulation on a Beowulf- cluster the results implied, that the Ethernet latencies –even on an Gigabit network– make it difficult to gain a decent speedup by adding more clusters. A solution was the use of special Myrinet network hardware, but the overall cost of such a cluster is high . We therefore concentrated on optimizing the single-CPU version of the queue simula- tion in recent years. To use cheap commodity hardware to speedup the simulation on a single computer would be of great benefit. In this paper several GPU based versions of this queue simulation algorithm will be presented to gain a relevant speedup on a single computer system.
External restrictions and specifications offer a framework in which carrier(s) have to operate. Customer demand (quantity, type and location to bring goods) is one very important specification for generating tours; e.g. time windows for delivery are one kind of restriction given by customers (see jsprit 2018). For our purposes we simply call this demand a “service”. Further constraints such as driving restrictions and tolls exist due to politics, either directly or indirectly because of court decisions. These constraints can be valid for certain areas, selected vehicle types, e.g. large trucks or vehicles with diesel engines, and/or certain times (see Cerwenka et al. 2007).
Abstract: Under the energy crisis and global warming, mass transportation becomes more important than before. The disadvantages of mass transportation, plus the high flexibility and efficiency of taxi and with the revolution of technology, electric-taxi is the better transportation choice for metropolis. On the other hand, among the many taxi service types, dial-a-ride (DAR) service system is the better way for passenger and taxi. However the electricity replenishing of electric-taxi is the biggest shortage and constraint for DAR operation system. In order to more effectively manage the electric-taxi DAR operation system and the lots of disadvantages of physical system and observe the behaviors and interactions of simulation system, multi-agentsimulation technique is the most suitable simulation technique. Finally, we use virtual data as the input of simulation system and analyze the simulation result. We successfully obtain two performance measures: average waiting time and service rate. Result shows the average waiting time is only 3.93 seconds and the service rate (total transport passenger number / total passenger number) is 37.073%. So these two performance measures can support us to make management decisions. The multiagent oriented model put forward in this article is the subject of an application intended in the long term to supervise the user information system of an urban transport network.
In this paper, we described a workflow for generating multi-agenttrafficsimulation scenarios based on OpenStreetMap. For detailed, calibrated studies, additional data will be necessary, but in principle, it is possible to set up a basic simulation based only on OSM and a commuter matrix. While it is still difficult to link public transport information to OSM, the ongoing efforts to establish a tagging scheme for transit routes is promising. Finally, using a network for the simulation which at the same time can be rendered as a high-quality background map has substantial benefits for the visualization of results. We think that Volunteered Geographic Information will play a large part in making it easier to set up and maintain large-scale trafficsimulation scenarios.
The problem of multi-robot exploration was tackled by Singh et al. to optimally plan the trajectories of multiple robots to explore a process . This process is modeled as a pre-learned Gaussian process and the optimal exploration problem is solved in a centralized manner. Decentralization brings advantages in terms of the algorithm’s robustness respect to agent failures. Chen et al. propose the use of Gaussian processes to monitor online trafficwith multiple robots in a decentralized manner . In , we proposed a decentralized multi-agent strategy to explore a magnetic field intensity in an indoor environment. We evaluated the algorithm’s performance with respect to several covariance functions and proved its convergence and scalability with simulations. However, both of the aforementioned works assume the model’s hyperparameters to be known. Ouyang et al. go one step further and propose an algorithm that also learns these as the process’ structure . They use this to actively sense an environmental phenomenon, which is modeled as a non-stationary Dirichlet process mixture of Gaussian processes. However, they validated the algorithm in simulations and did not tackle the challenges that ap- pear in an experiment. Similar to , in  the authors propose a decentralized algorithm with mobile robots that act as elements of a sensor network to monitor a physical phenomenon in a lab environment. However, they focus on the spatio-temporal monitoring; i.e. the sensor placements are fixed and the robots must decide where to move next in order to monitor the process. The main difference with our work lies on the fact that we consider all the positions in the environment as possible sensor placements. Then, we only aim to measure in some of the locations to reconstruct
In the study considered here, freight follows the same scoring (= utility) function as private transport, i.e. items (i) to (iii) above. More realistic freight-speciﬁc scoring functions need to be considered in future work. For the present study, the simulation is set up such that only re-routing is switched on as choice dimension, and freight activities are ﬁxed by their end times, which are the same as the departure times for the following trips. Thus, travel time savings are converted into score improvements by multiplying them with Eq. (2).
Abstract A spatial queue model in a multi-agentsimulation framework is extended by introducing a more realistic behaviour, i.e. backward travelling holes. Space cor- responding to a leaving vehicle is not available immediately on the upstream end of the link, instead the space travels backward with a constant speed. This space is named as ‘hole’. The resulting dynamics resembles Newell’s simplified kinematic wave model. Furthermore, fundamental diagrams from homogeneous and heteroge- neous traffic simulations are presented. The sensitivity of the presented approach is tested with the help of flow density contours.
Although the projected version of WGS 84 is widely used, it is also criticized by GIS experts for its lag of accuracy. The projection not only sacrifices the poles, but it is also distorts the original projection. Web Mercator uses mathematical formulas and parameters that make it incompatible to WGS 84 (National Geospatial Intelligence Agency 2014). The errors increase with a larger distance from the Equator and can reach an offset of up to 40,000 meters. Figure 2.6 shows an overlay map of the United Kingdom, where it becomes most visible how severe differences are. The South of the UK shows an offset of 33,000 meters, while the North shows 36,700 meters.
Traffic is highly influenced by network structure and human behaviour. Small changes in the human behaviour can lead to huge changes in the load of a traffic network. Current transportation models do not, and most of them can- not, research such random behaviour but always calculate a steady state. In our multi-agent transport simulation, we frequently observe seemingly ran- dom ”network breakdowns”, huge traffic jams that spread over a big part of the network, making a normal traffic flow impossible. This paper describes the investigations that were performed on the results of our large-scale multi- agent transport simulations in an attempt to contribute to the better under- standing of the dynamic processes in such simulations and, hopefully, better understanding and modelling of the real-world.
- 369 -
analysis (Schmoecker, Bell & Kurauchi, 2008). So they are commonly adopted for the strategic and long-term planning and evaluation.
It is difficult to make a global mathematical model to reflect the interactions between trains and passengers for its dynamic and complex. With the development of computer simulation, a growing number of simulation models have been applied to traffic analysis and evaluation. A micro simulation model has been proposed for road traffic to analysis the network through single vehicle moving and intersecting (Shang & Lu, 2006). Meignan and Simonin (2007) adopt a multi-agent approach to describe bus network as behaviors of numerous autonomous entities such as bus and travelers, through interaction between the main components to reproduce the passengers’ travel process of entire network, from passenger loads of network and waiting time to evaluate the operational efficiency. However, using simulation to study the dynamic passenger flow distribution for urban rail transit network is still lacking. In this paper, we will adopt a multi-agentsimulation model to describe the entire subway system, including the processes of passenger walking in station, waiting to get on and off, transferring and exiting, which can well describe the whole process of passengers travelling in network. In addition, this meets the computation efficiency for large-scale network.
Therefore, the majority of definitions contains two aspects: “a capacity issue”, e.g. “how many tourists can be accommodated without causing irreversible negative impacts on the destination”, and a “perception of capacity” issue, e.g. “how much tourism is acceptable before a decline in visitor and resident satisfaction occurs” (adapted by Coccossis and Mexa, 2004). As a consequence, the TCC should simultaneously focuses attention, on one hand, on the host destination impacts and population attitudes (Martin and Uysal, 1990) and on the other hand, on tourist satisfaction, two issues interfaced one with the other. In fact, the greater the intensity of tourist use and the level of saturation of the tourist assets are, the more limited becomes the appeal of the tourist attraction and the more intolerant become the residents. This happens mainly in the case of overcrowding, in mass tourist sites (Marzetti and Mosetti, 2008).
VRP, the great emphasis was put on heuristics accuracy and speed, whereas simplicity and flexibility were out of focus . As a consequence, the best state- of-the-art algorithms give very good results for many theoretical test instances of the static VRP, but they are hard to adapt to the dynamic real-world problems. Therefore, it is necessary to focus the future research on realistic VRPs. But this requires the development of a system that will be able to simulate various real-world vehicle routing problems thus allowing for both testing optimization algorithms and planning transport services. Such a realistic simulation has to incorporate realistically modelled dynamism of customer demand, traffic flow phenomena and fleet management operations. Especially the optimization of transport services in urban areas is extremely challenging due to high dynamics of traffic flow resulting in continuously changing travel times and often, depend- ing on the type of services, in high volatility of demand (e.g. taxi). Moreover, when considering city-logistics policies, many additional and often mutually con- flicting objectives appear as, for example, the reduction of the negative influence on the environment and on the local society, or the positive influence on city development.
The blue points in Fig. 1a are for the bike only simulation. In this case, ﬂow grows linearly with density, up to a value smaller than the maximum ﬂow capacity of the link, and then abruptly drops. The reason is as follows: Free speed determines the rate of the linear increase in ﬂow in the laminar regime. For a lower free speed, the point of maximum ﬂow is reached only at a higher density. In the case of bikes, free speed is so slow that the maximum ﬂow is not reached before the maximum density, so that the capacity regime does not exist and the laminar regime changes directly into the jammed regime. Similar plots for bike from survey data were obtained in past. 18 , 7 The latter also state that bike ﬂow rarely reaches high volume conditions.
Besides difficulties in specifying the requirements for a scenario like the one above, the construction of the system model can itself become a complicated and error-prone task. One obvious source for complexity that is inherent to any multi-agent system is the fact that there are many interactions between agents that may happen at the same time. If a microscopic view is applied as described above, this typically implies that the situations of individual agents have to be considered both for determining which interactions happen at a particular point in time and for calculating their outcomes. Furthermore, agents can also interact with their environment, which could include the physical world around them but also more abstract virtual entities that are shared between agents like a database or a network. This interaction with the environment can mean that agents react to events that originate from the environment. At the same time, actions performed by agents can change the state of the environment, which in turn might trigger events or influence other agents. It is obvious that a modeling approach that is suitable in such a setting has to provide means to capture a large number of effects and constraints whose scope ranges over the whole system. Doing this in an imperative programming language can lead to code that is very hard to comprehend and reason about. This can be experienced in typical simulation frameworks for general purpose languages like C/C++, Java, or Python that are designed for optimized performance. A viable alternative for describing the effects of interactions and events on the state of agents and the environment are rule-based languages where the outcome of a rule typically determines the value of one or more state variables for the next time step. In fact, the modeling languages that are employed by model checking tools are typically rule-based. However, for agents with more complex control logic, a purely rule-based representation of their behavior can be just as hard to follow. Altogether, the right solution appears to be a proper combination of several paradigms.
Agent based simulation (ABS) models have been developed quite far today. To describe the properties and power of ABS in the field of social science simulation, it is advisable to begin with its predecessor, the so called cellular automata (CA). A CA consists of a number of cells arranged in a regular grid. Every “agent” is a cell that is situated within other cells in a multidimensional array. In a two-dimensional setting this cell is located on a (seemingly) rectangular grid. In most CA models the border-cells are connected, so that a cell on the left “border” is connected to the cell on the “right bordered” cell in corresponding height of the grid. Analogously a cell on “top” of the grid is the neighbour of the corresponding cell on the “bottom”. So each cell has the same number of neighbouring cells and the seemingly rectangular grid represents the surface of a torus. In social simulations cells may represent individuals or collective actors like households, firms, communities or even countries. Each cell can be in one of a few states, for example, `on' or `off', or `alive' or `dead', represent attitudes (e.g. supporting one of several political parties), individual characteristics (e.g. ethnical origin), or actions (e.g. co-operating or not co-operating with others). Time advances through the simulation in discrete steps. After each time step, the state of each cell may change. The state of a cell at any time step is determined by a set of rules which specify how that state depends on the previous state of that cell and the states of the cell's immediate neighbours. The same rules are used to update the state of every cell in the grid. The model is therefore homogeneous with respect to the rules. Because the rules only make reference to the states of other cells in a cell's neighbourhood, cellular automata are best used to model situations where the interactions are local. CAs have been used as models in many areas of mathematics, biology and physical science, as well as social science.
order is announced. Thus, it is introduced in multi-agent system as the internal mechanism of an agent. The MNL model is generally based on random utility theory. This model is appropriate for evaluating some measures which include various behavioral cases on announcing evacuation orders or recommendations. Here, the MNL model consists of some kinds of information on mudslide disaster, phenomena of mudslide, communications with evacuators in their family, contact with neighbors, awareness of risk etc. These attributes and factors are composed of explanatory variables of the choice model. Thus, it is supposed that the case of executing the evacuation is 1 and the case of staying in home is 0. Then the common utility value of all respondents is provided by using the explanatory variables. The software used for the estimation of parameters in the model was LIMDEP 7.0 [Limdep Econometric Software, 1998].
The second requirement is highly related to the first point and deals with the prob- lem to establish a proper relation between the system and the real world it is sup- posed to model. A prominent example for the gap between research and reality are order dispatching systems for haulage companies: usually, the respective systems try to compute optimal routes and schedules for the companies trucks, but they fail to monitor the plan execution process and are thus not able to react to unforeseen situations such as mechanical failure of trucks or traffic jams making it impossi- ble to maintain the original schedule. An exception from this shortcomings is the
are aiming to solve problems for which the agent-based modeling approach is more suited than more traditional four step process modeling. The typical use-case starts with the collection and preparation of data. Then, several simulations are run on a cluster for several days and compared afterwards. Often, this must be repeated several times as data may contain errors that do not become visible until simulation results are retrieved. If no automatic calibration is available, repetition of simulation runs might be needed to calibrate the model. Multi-agentsimulation pays off when heterogeneous user preferences, time dependent user reactions, or/and microscopic modeling of pub- lic transport are studied, (e.g. Rieser et al., 2008; Grether et al., 2009b; Neumann and Nagel, 2010). At the time these studies were undertaken, there was no default soft- ware support for the required functionality; MATSim had to be extended by custom model components. Today, some of this functionality is available in MATSim or as a contributing project. However, modeling problems at this high level of detail make it hard to define a default methodology and implementation that suits every user. So, a user quickly finds himself in the second group, the researchers.
The typical method to couple activity-based demand generation (ABDG) and dynamic traffic assignment (DTA) is time-dependent origin– destination (O-D) matrices. With that coupling method, the individ- ual traveler’s information gets lost. Delays at one trip do not affect later trips. However, it is possible to retain the full agent information from the ABDG by writing out all agents’ plans, instead of the O-D matrix. A plan is a sequence of activities, connected by trips. Because that infor- mation typically is already available inside the ABDG, this is fairly easy to achieve. Multiagent simulation (MATSim) takes such plans as input. It iterates between the traffic ﬂow simulation (sometimes called network loading) and the behavioral modules. The currently implemented behav- ioral modules are route ﬁnding and time adjustment. Activity resequenc- ing or activity dropping are conceptually clear but not yet implemented. Such a system will react to a time-dependent toll by possibly rearrang- ing the complete day; in consequence, it goes far beyond DTA (which just does route adaptation). This paper reports on the status of the cur- rent Berlin implementation. The initial plans are taken from an ABDG, originally developed by Kutter; to the authors’ knowledge, this is the ﬁrst time traveler-based information (and not just O-D matrices) is taken from an ABDG and used in a MATSim. The simulation results are compared with real-world traffic counts from about 100 measurement stations.
ulation is modeled as agents with daily plans. Plans consist of activities with start and end times, and trips between activities that specify which mode of transport to use and which route to take. These plans are executed by the mobility simulation (see mobsim in figure 2). Traffic is modeled by spatial queues representing streets. Vehicles at the head of a queue can leave the link, when their earliest exist time (depending on the links free flow travel time) has past, link flow capacity is fulfilled, and enough space is left on the next link (i.e. spill back effects are modelled). With this, plans can get delayed which is evaluated by the scoring module. Some agents are allowed to adjust their plan, e.g. their route, departure time, or mode of transport (replanning). The co-evolutionary process of execution, evaluation and replanning is repeated iteratively to achieve a state where no agent has an incentive to unilaterally chance his plan – a (stochastic) user equilibrium.