Agent-Based Simulation - Science topic

I've just uploaded my introduction to Historyware that explains my research in digital history https://www.youtube.com/watch?v=DQQbeCGWMAg&ab_channel=AlexanderOsherenko

Kindly read this work may you can get a strong base about what you are looking for.

If you cannot download it, please let me know. I will try to download it from my side.

Regards

Is there any agent based software which can be coupled with COMSOL? and isn't COMSOL capable of solving agent based models itself?

You might want to start with lattice-boltzman schemes, as I think they offer an an easy to understand introduction into numerical simulations. There is this python library pylbm ( https://pylbm.readthedocs.io/en/latest/). You could check out their example gallery to see if it's sufficient for your purposes..

Davor Mance I really like the differentiation by Elinor Ostrom that you stated in the end. Do you, by any chance, have the reference for this distinction?

I think this always depends on the technology ecosystem one wants or has to develop in. Most popular BDI agent frameworks are Java-based (Jason, JACK). From an academic point of view, Jason is a natural top candidate, but from a practical software engineering perspective I would select based on the following criteria:

For web-based BDI agent development, let me shamelessly plug my own new library, JS-son: https://github.com/TimKam/JS-son (JavaScript-based).

The list will be helpful, especially if they are categorized by the type and application. If there is an overarching organization for simulation software (government or commercial), a link to that organization or main points of contact can be useful.

Please check these:

Please tell me if the attempt to open the zip file prospered. Regards

Hey Ameneh,

there is a whole host of papers out there that deals with the calculation of trust values by applying a trust metric on the experiences gathered by the agents. Depending on your concrete systems, the meaning of these trust values and the way to get to them can vary greatly. The seminal work in that area is Steve Marsh's Ph.D. thesis in which he discusses the most important issues. It was written 26 years ago, but it is still valid and I know that he is working on an updated version to be published as a book next year. You can find the thesis at https://dspace.stir.ac.uk/bitstream/1893/2010/1/Formalising%20trust%20as%20a%20computational%20concept.pdf

Based on this work, many people thought about specific trust models and how to make use of the generated values in decision procedures. The list of authors provided by Houcine is a good starting point, but of course there are many others. I am not aware of software solutions that you can use out of the box. This might be due to the fact that each system seems to have a very specific notion of trust that is not easily generalisable. The basic process is always the same: collect expectations about the behaviour of others, collect evidence about their actual behaviour, compare the two and derive a value through a metric, exchange this information (that would be the reputation part), and use the value in a decision procedure. But what the value means and how to get to it (the metric) differs greatly. The other problem is that there is very little standardisation in multi-agent system platforms, so you can't just have a package that can be universally applied.

Having said that, implementing an actual trust system yourself if you know your implementation platform, your metric, and the purpose of the trust values is not very difficult. Let me know if you need more pointers for that.

Best,

Jan-Philipp

1.Come up with a taxonomy/ hierarchy of heuristics which is simultaneously mapped to facets  of your problem/area of application. This is where multilevel decision making is addressed. The complexity of your problem dictates the complexity of the taxonomy, though not necessarily in a proportionate manner. As you could exploit the power of recursion to attain to different categories of the taxonomy.This will give you an idea of the kind of decision rules that are contextualized for your application. Recursive agent based simulation can be used to define or identify useful heuristics in the problem space.

2. Determine what kind of decisions are possible/ decision rules are applicable from the availability of data. Using principles from data-driven decision support systems/decision-making (DDDM).

3. An outline of possible decisions can then be inferred through a combination of 1&2. Point 1 is mainly generative. Point 2 focuses on implementation and constraints to the set of possible heuristics. May need to take a trial and error approach and reverse the sequence -point 2 then point 1, at multiple junctures in the design flow.

Obinna,

I would suggest that you look into publicly available sources such as Bloomberg APi's for obtaining data or yahoo finance (it has export functionality). What is important is that you get both the historical movement as well as volume of trading from whichever source that you target. Volume measurements will help you establish easy and reasonable points on which to measure trading exchanges as well as big movements(or minimal) in the market.

Based on this information and the strategy that you choose for each agent, then divide it into sell, no sell. What can be done afterwards is to let the agents do decisions on non tagged data and see if the predict the uptick or not and then quantify the loss.

If you have specific questions you can send me a direct message and we can work the details.

Regards

For example you could simulate humans as points

in the plane. Objects that have x,y coordinates.

In order to accelerate computation, store

references to these objects in an array where

each cell can hold a list of several humans.

Behavior: normally they avoid approaching

each other too close. Under normal circumstances

humans walk with a certain speed toward diverse

goals, avoiding collisions. In panic, they

try to reach one of several exit doors.

They try to follow distinct individuals

which appear to be on the right track.

Panic can appear in different levels which

leads to wrong decisions, whatever that

may be.

Literature:

Regards,

Joachim

I've looked at the Fishbone; and adored it as highly comprehensive. The only thing I really saw as missing was HR structure and process to recruit the right people for innovation/adoption. But I have a broad quibble as to whether the fishbone structure  isn't a generally applicable one, around technology adoption, rather than specific to Higher Education, an observation I feel highly relevant. The point is, in HE, I've seen ineffective technology as a barrier to adoption and I've seen resistance to change, but HE has a particular enabling factor for resistance to change, that you won't find elsewhere - Academic Freedom

First of all, just to mention that exist a great confusion on this because there is a great confussion regarding on what is a model and what is an implementation of a model.

To perform this classification is needed to be focused on the model and not in its implementation, that can be diverse.

Having said that, a DES model is a model where the EVENTS rules the dynamic behavior of the processes that  pictures our system understanding, the EVENTS are the elements that modifies the state's variables. Whit this idea, the models can be represented easily with DEVS, Petri Nets or Specification and Description Language (among other formal languages). In these formalisms is quite easy to detect that the EVENTS are the key element that rules the behavior, hence here we are talking of DES models.

In contraposition to this we can define a models where the EVENTS do not rule the dynamics, as an example, those models defined by Forrester diagrams (system dynamics) where a continuum modification of the state variables takes place. In that case the paradigm used to define the models is different (here we are not talking of DES models but continuous models).

MAS systems  (AB simulations) are those simulations where the entities (i.e. agents) can own a great complexity, can learn, and can modify his behavior depending on his knowledge and the information they obtain from the environment. The definition of this kind of models are based on the description of EVENTS that impact on the agents modifying its states. Hence MAS is an specific DES model where exists specific elements that are more complex, but, the dynamic modification of the states are mainly based on EVENTS, hence MAS are a DES models.

If you are interested on see a formal representation of a MAS system see the papers I attach. On they are clearly detailed why (although is not the purpose of the papers) a MAS model is a DES model.

De Kleer is a good start - some additional pointers:

Coordination for ATM

You may also wish to develop a variation on Chang et al.’s. (1993) research on distributed reasoning supporting multi-auditor cooperation. The auditors can be represented as agents. The Chang et al. presentation addresses the support for a team of cooperating auditors.

During the cooperation stage, the auditors must ensure that their default assumptions are not contradicted by the empirical evidence of other team members. The cooperation process requires a narrowing down of the areas of conflict, identifying the areas where further testing is required and the development of the explanation for the consensus opinion that emerges from the process (Chang et al.1993, p. 347).

As part of their analysis, auditors develop propositions or beliefs, based on their assumptions . These propositions may then be communicated to a more experienced auditor who may judge them as true, false or unknown (this triplet may be seen as a component of the linguistic). The judgments of the more experienced auditor are subsequently communicated back to the originating auditor who negotiates until a consensus is reached.

Chang, A., A. Bailey Jr. and A. Whinston (1993). "Multi-Auditor Cooperation: A Model of Distributed Reasoning". IEEE Transactions on Engineering Management, 20(4): 346-59.

[1]These assumptions are by nature default assumptions, which hold that in

   the absence of evidence to the contrary, the item under review is sound.

Did know JaCaMo?  Take a look:

Hi James!

I would say that the concept of (social) power is basically linked to the idea of the ability to control the flow of "scarce" ressources in a given network. Hence, persons occupying strategically important positions - sometimes also called "brokers" (R. Burt refers to "structural holes" as the structural complement; see also his book on "brokerage and closure") might exercise power. Such actors possess instrumental social capital; they provide scarce ressources to others and receive obedience in return.

I published an article in Connections (that you can find on reserach gate) on a way to concepzualize different forms of social capital and, hence, to identify powerful broker roles in social networks based on a distinct concept of "betweenness":

"Measuring the Social Capital of Brokerage Roles"

Kind regrads, Volker

You should have a look to teh journal Economic Dynamics and Control, with a special attention to Giovanni Dosi's articles.

Regards,

MV

There is a new standard for designing and reporting ABMs. A group of modelers that includes two leading authors of ABM books, Volker Grimm and Steven Railsback, published the ODD Protocol, then followed up with an update in 2010 in the journal Ecological Modelling.

doi:10.1016/j.ecolmodel.2010.08.019 

@Roman and Bernardo: to directly answer the question, there's an overview of ABM micro/macro models in an upcoming Handbook chapter (see attached):

"Section 2. Recent Developments in Agent-Based Macroeconomic Modelling"

@Roman: Concerning complexity, I have the feeling that the Eurace@Unibi Model is perceived to be more complex than it actually is. (Disclaimer: I have a personal interest in the model.) Yes, there's a lot of details, and yes, there's a lot of code. But it also models 5 different types of markets! At least we tried really hard to make everything open, and you can now also download the entire code and run it in a virtual machine, something not many models have allowed you to do so far, in fact.

Evolution is a process that has no goal or aim. It's not a teleological process.

When we run a simulation with a determined goal, it's our aim, not evolution's that is being pursued.

Aleksandar suggestion is confusing an a posteriori observation with an a priori goal. It is like saying that the goal of evolution is produce biodiversity. Simply wrong.

Check the work of Susan Spierre Clark. She has done some interesting game design similar to ABM but not quite ABM.

This really depends on what you are after. From a technical perspective, you have to determine if your system allows you to scale you model. Some tools work well for a few hundreds agents but will fail to scale for thousands. Can the tool be parallelized for use on clusters/clouds.

What language do you need to use to express your ABMS will determine the availability of other tools, such as graph or statistics libraries. You certainly want to be able to use efficient data structures as complex social systems often are very large.

Another aspect is that of data representation and visualization. Does the toolkit allow you to easily interpret the results? What visualization capabilities do you have available? Can you generate file formats that easily integrate into existing visualization systems?

Main Advantage is following – GA, as all other random-search oriented optimization algorithms (Simulated Annealing, Cross-Entropy, etc.) doesn't require any information about structure of the function to be optimized and uses it as Black Box. Classical optimization methods should use some information. Consider, for example, function FMINS from MATLAB. To search minimum of the Rozenbrock function of the two variables, it should use initial point. If initial point is "good" (e.g., [-1; 1]) - will be success for optimization. But if you don't know, how to select appropriate initial point, search will be wrong with very large error. Random-search oriented optimization algorithms don't require information about initial point and successfully find optimal solution for Rozenbrock function even for 10 variables.

NetLogo look very good for beginners, but this is "dead end".

It have unique, strange language and biased philosophy of model design. From any JAVA based environment you can switch quite easily to any other based on JAVA, C# or even C++ or CUDA (when you need higher efficiency and scalability), from NetLogo you have to learn almost from scratch.

Part of my doctoral reseach on NPCs and PCs are presented as game design patterns: Lankoski, 2010, Character-driven game design. https://www.taik.fi/kirjakauppa/images/4463b7dc09f925da01eaf73cbc72eed9.pdf

Patterns are in Appendix 2 (pages 61-74). More detailed analyses are presented on pages 116-155 & 162-181.

Staffan Björk has been extending some of these patterns in Gameplay design patterns 2.0 wiki http://gdp2.tii.se/index.php/Main_Page

My only direct knowledge of Holland's DGS' is from this 1.5-page review by Adami:

http://www.sciencemag.org/content/338/6113/1421.full.pdf

Are you applying the DGS to security? ... Or better yet, what aspect of security are you applying it to ? The problems of, say, spear phishing are likely to be very different from those of, say, airport security -- would you agree?

actually liveability is one of the main elements which indicate a good city. So if it consider in the first step of an urban planning, I think that lead to a good urban designing and also architecture. it highly depends on the act of the city in its regional and its attraction.

You can check the work of Giuseppe Vizzari from Milan:

He works on agent-based simulations of crowded places, in particular he exploits agents to simulate people behaviours.

Hope this helps,

Giacomo

You could try some of the generic discrete event simulation tools to model this (such as arena, plant simulation or extend)

Very interesting! Swarm particle optimization method is often used to solve the problem of finding the global optimum of a dimensional function. The basis of method are the swarm intelligence algorithms, which concern with the design of intelligent multi-agent systems based on the collective behavior of insects (ants, termites, bees, and wasps) or other animal societies (flocks of birds and schools of fish).

Below I am giving some references:

[1] H. Schmeck and J. Branke, Designing evolutionary algorithms for dynamic optimization problems. Theory and application of evoluacionary computation: recent trends, S. Tsutsui and A. Ghosh, Eds., 2002, pp. 239-262.

[2] J. Kennedy, R. Eberhart and Y. Shi, Swarm Intelligence, Los Altos, CA: Morgan Kaufmann, 2001.

[3] K. Parsopoulos and M. Vrahatis, "Recent approaches to global optimization problems through particle swarm optimization," Natural Computing, pp. 235-306, 2002.

[4] T. Blackwell, R. Poli and J. Kennedy, "Particle swarm optimization. An Overview," Swarm Intell., vol. 1, p. 33–57, 2007.

There exist many technical aspects of the algorithms, which recently I have been interested too in a different area of application. So if you have any questions we can discuss it.

The GLOWA Danube project is also a real example of use of agent-based models by decision-makers.

Here are some links to the project:

Dear Peer-Olaf,

I think we might have met in EASSS 2009 in Turin... but maybe I'm wrong.

Anyway, from my perspective, it all comes down to your first question... and the answer is probably "because most people actively working in this area do not really come from Computer Science"...

From my understanding, moreover, it is relatively simple to use a class diagram to capture static aspects of a domain, which in this context would mean to represent agents, relevant aggregated abstractions, relevant 'objects'... it would also help giving an explicit account of additional "things" (in a better English abstractions) you put in the model to manage some mechanisms... but you'd need additional diagrams to capture dynamic aspects of the model, and this would make things more complex (you probably need to be quite good at UML modeling to really capture all the necessary aspects)... and frankly I think it would be difficult to achieve the level of formalism you can get with equations.

I guess you might be interested, on one hand, in all the systematic efforts carried out in the AOSE community about meta-models and modeling methodologies. On the other hand, I think that the idea you propose of defining a systematic way of translating equations into agents is original, and you might actually exploit work carried out in the AOSE community.

Sincerely

Giuseppe

If you want to know the whole story about tsunami related evacuation simulation with matsim you may look into my dissertation:

or the newer work by Christoph Dobler:

For the multi-modal evacuation you can check this tutorial:

If you planning to develop a hybrid model like Anh et al. at al did, the work of Bourghout might also interesting for you (see, e.g.,

Burghout, W. and J. Wahlstedt, Hybrid Traffic Simulation with Adaptive Signal Control. Transportation Research Record, Vol. 1999, 2007, pp. 191–197.)

The highest IP score rests with Molecular Systems Biology. It is now back to eleven, probably since they started including reviews as well. From you title you may however fare better with a journal that is a bit more more open to methods-focused papers. BMC Systems Biology might be a bit lower in impact but is well read in the field. Submitting to Bioinformatics (5.3) might also be an option if your article fits into their strict page limits/layouts. PloS Comp Biol is at 4.9 now and is also a good target. If your study contains experimental/clinical data that is linked to your modelling definitely try to give those a go. If it is a purely theoretical study it will definitely be more difficult.

This may help "Comparing Cartel Behavior: A Simulation Analysis with

the System of Cartel Markers (SCM)"

We have just started looking at the topic of "Test Driven Object Oriented Simulation Modelling" and are planning to extend this to "Test Driven Agent-Based Simulation Modelling" in the near future.

Here is a list of papers we found useful for our studies:

- Sawyer and Brann (2009) How to test your models more effectively - Applying agile and automated techniques to simulation testing

- Dietrich et al (2010) Validating UML simulation models with model-level unit tests

- Nguyen et al (2011) Testing in multi-agent systems

- Djanatliev et al (2011) Veritas - A versatile modeling environment for test-driven agile simulation

- Coyne et al (2008) A methodology for unit testing actors in proprietary discrete event based simulations

- Pitt-Francis et al (2009) Chaste - A test-driven approach to software development for biological modelling

- Guercan et al (2011) Towards a generic testing framework for agent-based simulation models

- Collier et al (2007) Test-driven simulation development using Repast Simphony

- Kleijnen (1995) Verification and validation of simulation models

- Kleijnen et al (1998) Validation of trace-driven simulation models - A novel regression test

I´ve had the same feeling just by doing some preliminary analysis of trends in scientific databases (Scopus, WoK), were one is going down, the other is coming up both in publication rate and in citations. However, I am not so sure about industry yet.

Thanks a lot for answer, but a bit of elaboration on the same would have helped a lot

Dear Samuel, I think that the NetLogo is a very good modelling tool for the biology. Maybe, it looks like a toy :-), but it is possible to create useful MAS models. Try to look on the wepage http://ccl.northwestern.edu/netlogo/models/ containing models created in the NetLogo. The next alternative can be Repast. It is free as a NetLogo. It offers different programming languages: Groovy, Java, state-charts and Relogo (similar syntax as in NetLogo). The GAMA generic multi-agent environment can be also useful (http://code.google.com/p/gama-platform/). There are good tutorials and examples. It uses quite untypical language GAML, but I think that it is not hard for learn as in case of NetLogo. The epidemiological model SIR was created with the aid of GAMA (for example). I hope that these advices will help to you :-). Best regards, Martina Husáková

Hi Patrick,

Many thanks for your answer. I am working on an algorithm for dynamically building an organizational structure in MAS. I am looking for a proper test bed or application to assess the algorithm I propose and probably compare it with the algorithms alike. The applications that you suggested would be good matches; however, they have their own complexities that I don't like to involve with, at least at this stage of my research. I wonder whether there is any simple handy problem.

I currently work on the subject, I inforler yours.

@all : what great answers ! Thanks so much.

@ Franziska : traffic simulation could be a good idea (and we have the infrastructure to do that), but I'm always afraid, in this kind of models, about the amount of details in which you have to dive to obtain something satisfying. Anyway, your pointer to the work of Dornhaus is just perfect (didn't know her work).

@ Arend : well, I did not think about evolutionary models, but you hit a point there. This is indeed a great way of showing the interest of ABM vs. other techniques.

@ Stefan : hope the other answers will contribute to your lecture as well :) Daisy World is also a nice model I totally overlooked, probably because I was focused on "real" problems. But, in a sense, since I have six groups of students, it could be a good idea to introduce one "abstract" model in case…

@ Desi : one other work I did not know. And a great discovery. Thanks !

@ Derek : I totally agree with the — highly pedagogical — idea of building models of incremental complexity. That's what we use to build our tutorials for GAMA, for instance. My only concern with this course is that I have at most 4 weeks left with the students, which makes it quite difficult to implement a nice progression (for instance, if I want them to present their intermediary findings or implementations). However, I might very well reuse your two first models, given the fact that I already worked with them on a "simple" Schelling model of segregation.

Meanwhile, I have also worked a bit (and not only waited for your answers :) ), and found two interesting ideas :

- the different papers of Charlotte Hemelrijk about the DomWorld models (dominance hierarchies among primates). The ABMs (or IBMs) are more or less given in them, but since each paper deals with a different trait or behavior, the interesting question is how to couple them in order to obtain a quite complex artificial society.

- This (new) paper : http://www.plosone.org/article/info:doi/10.1371/journal.pone.0044501?imageURI=info:doi/10.1371/journal.pone.0044501.t001, about the importance of the "negative feedback" in ant foraging. Combined with some already classic toy ant models, it could lead to some nice models.

Plus things I already had under the hood (water flow modeling like the one depicted in the RIVAGE project, plant hoppers invasions modeling based on the life-cycle of individual insects, and a few other ones).

By the way, while looking for all this, I thought it would be great, somehow, to maintain this type of resources somewhere (as I'm probably not the only one scratching his head to find new and not so obvious models to implement with students). I know the domain quite well, but I was very surprised to see how difficult it was to find something else than simple toy models. And sites like openABM, for example, are not of a great help in that respect. Any volunteer :) ?