What do you mean by seamless integration? They are different approaches, MAS focusing on individual behaviors and SM on the system as a whole.

The software NetLogo allows you to model both MAS and System Dynamics, maybe it can be useful for you.

I also attach a paper explaining how to convert an Equation-Based model to a Agent-Based model. System modelling is not EBM, but this can help to find the seamless integration between both approaches.

The models based on SD are aggregate models which basically mean that you are dealing with different conceptual entities. From the SD perspective you are not working with individual entities as agents but with stocks where agents cannot be distinguished one from another. The system dynamics is a top down approach where you must think globally about your system in terms of stocks, reinforcing and balancing loops. Under a mathematical framework a SD model is equivalent to a system of ordinary differential equations.

On the other hand, Agent-based or Individual-based models are built bottom up and you think in terms of rules describing how an individual agent changes their internal state as function of the state of other agents, the environment and so on. In that case the modeling task is the understanding of the local structure the system under study.

Perhaps the most useful kind of integration of the two worlds is using the tools of system dynamics (i.e. feedback and reinforcing loops) when trying to describe the rules governing the agent behavior and the state variable update, this will lead to a much more structurally solid model.

By seamless integration I meant an approach wherein ABM is used to construct the aggregates and once the aggregates are constructed SD modeling takes over. It is bottom up approach. The reverse (Top down) approach is also possible. I am looking for an example of integration through either of approaches- preferably the bottom up approach if anyone has done that. 

SD is based on a system of equations that governs the dynamics of macro-level variables. Because of that, it might be difficult to start with an SD model and add the ABM layer. ABM works at micro-level, governing the behaviour of agents. That behaviour, in part, depends on the values of macro-level parameters. Some of these macro-level parameters are direct aggregates from agents' characteristics, while others are either set externally or are more elaborate functions from agents' characteristics and/or other macro-level parameters. These functions might well be borrowed from SD. In practice, that means that in your agent-based model, you'll have another class (or set of classes) that governs the dynamics of macro-level parameters according to the equations that stand behind the SD model.

I agree with Alexander Tarvid. Basically, what you are looking for is an agent-based model where part (or all) of your rules and parameters affecting agents' behaviours and states is modeled as SD. This could be a good choice to model non-linearity, loops and feedback about aggregates of agents (i.e., at the group level); the only doubt I have concerns the mathematical tractability of your results. As you know, agent-based modelling and mathematical models (e.g., SD) are conceptually opposite in the basic meaning that the first is the archetype of bottom-up approach (as already others said before me), and the second ones are those of top-down. ABM are affected by stochasticity of initial conditions and allow us to make group level behaviour emerge as result of micro-level interactions (i.e., agents); mathematical models allow us to perform rigorous a-priori analyses and to better "quantify values and measures". So, I am not sure that combining ABM and SD as you suggested will allow you to keep the best from both approaches... Hope this make help

System dynamics can also be microscopic. See my publications.

I agree with much that is in the answers already given in response to the RAI’s question. My view, from a classical AI perspective, is that SD tries to provide a useful model in terms of feedback loops etc which is a kind of summary or high level simplification of what would be a massively more detailed model were it created at the ABM level. This higher-level simplification may work well, but will only do so within a limited range of overall behaviour. For some uses of the model, it will have to be fully detailed down to the agent-based level. This is especially so if the thinking and decision making of human individuals has a significant impact.

I have myself implemented on a computer some aspects of the above simplification process -- see my little (and terse!) paper “Algorithmic Analysis of Production Systems…” -- readable on RG and downloadable – which does report some relevant algorithms (implemented and to a degree tested – for convenience using a particular type of programming language confusing called “production systems”) but arguably further work is needed to develop the ideas. The algorithms described look for ANY simple models starting from an ABM base, not necessarily for SD type models. So maybe that is a little more “seamless”. Details are in the article.

Jim

The difference between SD and ABM is not about scale, but about determinism. In a SD the components must follow the behaviour ruled by their equations (include probabilistic, chaotic or any other thing that you wish). You need agents when you can not model the system by a set of equations.

Hybrid multiagent architectures work as you describe (somehow): you have a first response, using the SD dynamics, and this response is refined with some intelligent behaviour. This ensures the system to have at least a response available. Try, for example http://abcresearch.org/papers/ESOA06Hybrid.pdf

You can find an example frome me here. I integrated individual purchase decisions of agents in a system dynamics model of the automotive market. 

http://pubsonline.informs.org/doi/abs/10.1287/trsc.2014.0526.

Thanks for the article.