environment and different decisions relating to different time horizons can be mapped. Liedtke (2009), for example, develops such a model for Germany. He explicitly models the decision of two main agent- types: The shipper and the carrier. Shippers can decide about shipment size and carrier choice. Carriers construct truck tours with a vehicle routing heuristic. Both iteratively interact with each other in a market environment and make experiences from past iterations. Roorda et al. (2010) set up a conceptual framework for agent-based modelling of logistic services. They identify a number of agents, their respective behaviour and important facilities in the freight system. The agents coordinate by means of contracts. The contracts are a result of market interactions. Shipper-Carrier relations are set up by logistic contracts. Given those logistic contracts the carrier conducts a number of logistic decisions to fulfil them. First, the carrier decides about the transportation mode. The possibilities include using only trucks as well as intermodal combinations of truck, rail and marine. Secondly, for each of those transport modes – in the following we name this transport chain – further consolidation decisions are conducted. That is, for each leg in the transport chain, vehicle type choice, vehicle scheduling and route choice is made. Ramstedt (2008) design a multi-agent based simulation of transport chains. They identify the transport chain coordinator (TCC), the transport buyer (TB) as well as the transport planner (TP) to be key decision makers on the transport side. The TCC is the interface between product demand, production and transportation choice and matching product suppliers with transport service providers. The TB manages the transport chain and its corresponding legs. The TP is the carrier actually owning a vehicle fleet and conducting the physical movement. Thus, transport chain choice and carrier choice are explicitly modelled.
Maintenance of the implementation was quite cheap during the last years. Fig. 3a shows in green the changes recorded by subversion to all code within the package org.matsim between 2010 and 2012. Changes to the package org.matsim.signalsystems, which contains the model speciﬁcation and the ﬁxed-time control implementation, are shown in red. While org.matsim in total was changed quite fre- quently, changes to org.matsim.signalsystems are rare. This means that for all classes within org. matsim.signalsystems, coupling is low. The changes to the signal systems package occurred mainly due to feature improvements and bug ﬁxes by the developer responsible for the extension. E.g. during the period 2010–2012 a feature for visualization was added to the package, also the intergreen logic was developed. Fig. 3b shows same information as Fig. 3a, but removes all changes due to this bug ﬁxes and improvements. Red changes nearly disappear, reinforcing the argument that coupling is low.
an alternative. Slow speeds of the alternative mode implicate a dominance of the air transport mode. If there is a seat on a flight, travelers receive a higher score than by traveling on the alternative mode. For the first 500 iterations, some virtual persons try different routes, times, or modes. Thus, during this phase a certain amount of seats is not occupied. More virtual persons validate the air mode as good choice than there are available seats. The corresponding plans get a higher score than the plans for the alter- native mode and shift towards earlier departure times. Recall, that the plan with the lowest score is deleted when the number of plans excesses a certain threshold (Sec. 2.1). For the first 500 iterations, this is the plan for the alternative mode. If plans for the alternative mode are recreated, they are copied from an air mode plan and obtain its departure time. Then, choice dimensions are switched off. Only the logit model is used for selection of existing plans. Travelers, that have tried the alternative mode in the previous iteration, switch back to the air transport mode with a high probability. This results in a lack of seats that are then allocated by time of arrival. Passengers rejected to board their flight get stuck. The probability to avoid getting stuck is higher the earlier one arrives at the airport. As the plan for the alternative mode is deleted before air transport plans, in most cases the plan database of a virtual person no longer contains this option. Otherwise, passengers may switch to the alternative mode with a tendency towards an early departure.
From animal experiments it is known that a certain dose rate of a cytotoxic therapeutic agent will lead to the cell death of a relatively constant fraction of cancer cells. Normally 99.99 per- cent of the tumor cells will undergo cell death. For a tumor with about 10 11 cells this means that 10 7 cells will survive the treatment. Therefore, the tumor should be treated as early as possible. However, most of the cancer diseases cannot be diagnosed before the tumor has reached a certain size. Assuming the development of a tumor from a single, malignant cell and an exponential growth behavior of the cell, the tumor reaches two centimeters in size after 30 duplications (10 9 cells). Depending on the type of cancer, this can take one day, two weeks, several months, or more than one year. A tumor consisting of 10 9 cells needs only ten further duplications to reach the critical tumor mass of 10 12 cells. Now the tumor has a diameter of 20 cm and this leads to the death of the patient in most cases. Three quarters of the time, from the cancer development to the death of the patient, the tumor will not be diagnosed (Dingermann et al. (2002) ). There is only a very limited time frame for a successful cancer therapy. The large majority of today’s cancer therapies are based on the removal of solid tumor masses by surgery, and a plethora of physical and chemical treat- ments like chemo- and radiotherapy, that induce the death of all particularly sensitive or rapidly growing cells. These approaches are variously combined in order to optimize the therapeutic efficiency. It is common knowledge that today’s cancer therapies have serious side effects. One reason for this is that neither cancer cells nor the cancer cause are directly targeted. A new approach in cancer therapy is the development of a treatment that targets the cause of a cancer. The TRAIL based apoptogenic therapy seems to be a great promise in tumor targeting therapy. It combines high cell killing potential with tumor cell selectivity.
The e-toll scheme is based on base rates of ZAR 0.70/km, 2.10/km, and 4.20/km for Classes A, B, and C, re- spectively. A discount of 25% for e-tag holders is o ﬀered. For commuters within the province we assume an e-tag penetration of 40%; for bus, 50%; paratransit, 40%; commercial vehicles 60%; and external commuters, 25%. Period discounts are speciﬁed on an hour-by-hour basis, and range between 0% and 20% for a weekday. A public transport discount is o ﬀered to both buses (55%) and paratransit (30%). Although a frequency discount is also oﬀered, it was not considered in this paper since the current agent-based model only represents a single full day. As stated earlier, toll is calculated whenever an agent enters a tolled link in the network. Both the time of day and the agent ID are known, so one is able to accurately determine the correct base rate. The relevant discounts can then be calculated and applied to the base rate.
At any rate, the comparison of Fig. 5 to Fig. 6 makes clear that the selection of the aggregation procedure can lead to quite different equity interpretations: While the simple but conceptually problematic summation of utilities leads to similar gains across all income groups (Fig. 5), the willingness-to-pay, sorted by income group, implies that high income groups would have a disproportionately high willingness- to-pay for the measure considered (here the pt speed increase; Fig. 6). This is, in fact, quite intuitive: Higher income groups have a disproportionately high willing- ness to pay for good schools, a good health system or a good transport system. In this sense, a progressive tax may not even be re-distributive with respect to such types of government expenses, since it just reflects the willingness-to-pay for improving the corresponding services. 16 Let us stress once more that these different attempts of measuring welfare rely on exactly the same description of human behavior. This highlights that the model predicting the system’s reaction to a policy change is inde- pendent from the different interpretations of measuring welfare.
Although it is not a technical restriction, SALMA is focused on modeling and simulating multi-agent systems. Multi-agent simulation has been adapted in many different fields, which has resulted in a broad spectrum of more or less specialized approaches. Widely used examples for domain-independent frameworks in that area are MASON (Java) [LCRP + 05] and RePast (Java, C++)[Col03]. Software packages like that offer highly flexible APIs at a rel- atively low level of abstraction. On the other hand, there are modeling and simulation approaches that are specialized on particular applications, e.g. the MatSim framework for multi-agenttransport simulations [HNA16]. SALMA tries to provide as much flexibility as possible with regard to fitting simula- tions to the characteristics of the modeled domain. Most of all, this includes SALMA’s ability to vary the level of detail and abstraction of the system model within a broad spectrum, which was discussed thoroughly in this chap- ter. However, the logic-based generic representation in SALMA is inherently much more computationally expensive than optimized specialized approaches. In particular, the application of SALMA might not be practical for models with very large numbers of agents, which is typical, for instance, in more re- alistic traffic simulation experiments. In such cases, it could still be beneficial to use SALMA as a supporting approach for analyzing certain parts or mecha- nisms of the model from a microscopic perspective. This will become even more apparent when SALMA’s abilities for statistical model checking are discussed in the next chapter.
Figure 5: spatial distribution of vacant dwellings 5 Conclusions
In this paper, we have reported some results of the current implementation of the agent- based systems of land use dynamics, titled PUMA. Due to space limitations, we could only very briefly summarize the motivation underlying the system and the kind of models that are used. It is important to note, however, that the process of developing the system is to specify the scope and architecture, exploring the system first on the basis of easy to implement, well-know models and gradually replacing these with richer, new, behavioural models. Agents to be included in the future are for example cognitive agents capable of activity-scheduling and rescheduling behaviour, learning the environment and capable of adjusting their behaviour, agents for simulating housing search and choice, incorporating negotiation between developers and potential buyers in a dynamic context, and agents simulating life trajectories and their impact on transport decisions. These models have already been conceptualised and their implementation is now being tested.
2. Theoretical background
There have been several spatial ABMs, which usually lack of a sound economic background, developed to demonstrate how the individual behavior of agents with bounded rationality and partial information can lead (or approximate) system-optimal equilibrium patterns, similar to analytical optimization models. Fan et al. (2000) presented a new economic geography (NEG) model with endogenous land use, labor mobility, inter-industry purchases and N-locations in one- or two-dimensional space, to underpin the development of a general equilibrium model of urban systems. Sasaki and Box (2003) showed how an optimal global spatial formation of land uses, that of von Thünen’s rings, emerges from simple agents acting on local criteria, without any systematic optimization functions at the system level. Recently, Heikkila and Wang (2009) indicated how an ABM version of the Fujita and Ogawa (F-O) model of household and firm location decisions across interdependent and spatially differentiated markets for land and labor in an urban setting yields equilibrium land- use patterns which are fully comparable to those described by the analytical F -O model.
A partial way out is offered by the approach chosen within METROPOLIS (de Palma/ Marchal 2002), which selects departure times of trips based on desired arrival times and schedule delay penalties. Given a time-dependent toll, travellers can react by selecting new departure times. A remaining problem is, however, the fact that trips and in con- sequence, decisions are not related to demographics. In addition, every model that uses single trips will have problems predicting useful reactions of travellers that span the whole day. This is because trips in real life are embedded in a complete day plan and are not meaningful just as stand-alone trips. Trips lead people from one activity to an- other, and in most cases the activities have a higher importance in the daily schedule than the trips do: Stores have opening and closing times, work places have fixed times when one has to be present, a full-time employee has to work about eight hours a day. This means that travellers cannot escape a toll at will, but have to trade off between different utilities (working eight hours, being at a shop when it has opened, etc.) and disutilities (paying a toll, being late for work, etc.). Thus, a toll may influence the complete daily schedule of a person, and not only the period when the toll is charged.
During the course of this thesis, a multi-agent system capable of automated meet- ing scheduling will be designed and subsequently evaluated by means of a proto- type implementation. The system’s design will have a strong focus on behavioral flexibility, meaning that core modules implementing the agents’ behaviors will be independently replaceable, thus resulting in a design that is both apt for further re- search and implementations destined for the end-user. As a multi-agent system, it will be capable of carrying out its functionality in a distributed manner without the requirement of any centralized entity. For such a system to be devised, the common approach to meeting scheduling, as it is carried out without the aid of automated systems, must be analyzed, with the objective of gathering the main requirements for successful automated meeting scheduling.
The coarse design diagram holds already the organizational structure for the multi- agent application. Thus the tasks can be directly derived. There will be n role modeling tasks m interaction modeling tasks and o ontology related tasks. All tasks can then be approached concurrently. This is sketched in Figure 9.1 . Note that all role, interaction and ontology tasks types are independent (concurrent) from other tasks of types roles, interaction and ontology. However, the concurrency is restricted, which is indicated by the dotted lines between interactions, roles and ontology tasks. This means that the implementing developers have to agree on a common interface, e.g. for the roles that participate in an interaction. During integration the independently developed system artifacts are assembled, their inter-connectivity tested and possible errors found, located and fixed (debugged). The outcome of the integration are milestones of prototypes that allow the developers – together with the experience gained during integration – to reconsider their previous design. We conceive the process iterative within each task type as well as over the process (indicated by the backwards arcs in the Petri net).
The concept of time also divides planning approaches into several classes. Whereas plans with a serial execution concept are called linear, there are application domains, such as multi-agent domains (see Section 10), in which it is convenient to have nonlinear plans (see Figure 4) with concurrent or even parallel actions. Nonlin- earity is obviously more general than linearity and, as we will explain in Section 3, it can be modelled by incomplete knowledge about the temporal ordering of events. A further advantage of nonlinear to linear planning is the reduced search space of plans. Each nonlinear plan is a representative of its linearisations (completion to a total, temporal order) and the planning process tries to keep the number of temporal restrictions as small as possible. The opportunities for conicts between nonlinear branches, however, are increased. It depends on the appropriate planning domain whether a linear or a nonlinear approach will be the better choice.
V n (U) + Φ(X (...U...))/µ of its individual utility fun
tion and the log-likelihood. The weighting parameter µ determines the importan
e of the behavioral prior information represented by the original utility per
eption. If µ is
hosen very large, the likelihood term vanishes and the agent a
ts in a way that is fully
3 Simulation results
Following the simple model presented in chapter three of Godley and Lavoie (2008), at the beginning of the simulation money is created through public deficit covered with the emission of bonds bought by central bank. Thus, in the first period of the simulation income is zero. Public spending, in the form of transfers to the household, is partially used to fund the formation of firms and banks. Thanks to public funds emissions the artificial economy gradually starts to reproduce simple economic processes: firms employ workers, pay wages, produce goods and try to increase their productivity. Banks funds private and public debt. Workers pay income taxes, save part of their disposable income and consume goods produced both domestically and abroad.
is queried from the EU-DEM model. In conjunction with the link length, the average link slope can be computed. This information is stored in the object attributes file. Based on information of OSM, values of the other additional attributes like highway type, cycleway type, and surface are stored in that file as well. Using highway type, surface, and slope, the free speeds for bicycles of each link are computed as follows: First, the minimum of the speeds according to the highway type and surface as outlined in Tables 1 and 2 are determined. This speed is then increased or decreased by a factor depending on the link’s slope. Ensuring that a link speed is never assigned lower than a defined minimum of 4km/h, the resulting speed is stored in the network file. Via relatable identifiers, information from the network file and the attributes file can be matched for the computation of the utility function.
Each person in the synthetic population obtains a second plan that uses the alternative mode. With this population the simulation is again run for 600 iterations. Like in the previous simulations 10 % of the virtual persons may shift their departure times while another 10 % seek a different route between origin and destination in the air transport network. Additionally, further 10 % of virtual persons may change mode, i.e. they can switch between the air traffic mode or the alternative mode. After 500 iterations all choice modules are switched off, thus for the last 100 iterations the logit model is used by by passengers to select one of their plans. From the output of the 600th iteration the same numbers as for the previous simulation runs are calculated and depicted in Tab. 2. If the speed of the alternative mode is 100 or 150 km/h the mean square error is quite similar to the previous results while the mean relative error is even less. The number of stuck passengers however is remarkable reduced from approx. 1500 to 185 or even 69. Alternative mode speeds higher than 150 km/h further reduce the number of stuck passengers while the relative error is quite similar. In contrast, the mean square error is increasing the higher the speed for the alternative mode is set.