There are a variety of modelling paradigms that can be used to represent and simulate real-world systems, but each comes with its own benefits and drawbacks.
Effectively choosing between modelling paradigms is a question of minimising bad trade-offs in the model creation, verification, and validation process.
We use a variety of modelling paradigms, each best-suited to a different set of problems, to enable speedy, productive systems simulation:
- System dynamics (SD) models take a global and systematic view of how dependencies and flows within households, organisations, and societies.
- Discrete event (DE) models can be used to monitor processes and sequences, often in industrial, manufacturing and supply-chain contexts.
- Agent-based models (ABMs) are employed to capture the ‘emergent’ impacts of entities interacting within systems, the results of feedback loops, and the consequences of adaptation in complex systems where entities are able to respond to their environments.
But single models may be created that combine multiple paradigms, and the open-source software we’re building at HASH specifically seeks to address this challenge.
Integrating abstract paradigms within agent-based models
Agents in ABMs need not always represent individual people, and can represent a wider array of actors (e.g. households, firms, or even nation-states and entire species).
Provided the entity in question is something with agency, it can be an agent — and as a result, it’s common for whole organisations to be classified as ‘agents’ when it comes to building ABMs.
A human resources department modelling employee morale levels, for example, might choose to model their entire workforce as a single agent for the purposes of finding out whether the activities of another organisation (such as a team-building firm or pensions provider) had any effect, then also use a discrete events approach to model the process of employees conversing with mentors and how that affected their morale.
Another common permutation is the simultaneous event-agent approach. If you are analysing an assembly line, for example, it can useful to see each personnel member at each station as both an agent (in order to measure how and by what processes they are affected as an individual staff member) and as a crucial component of a discrete event sequence (in order to see where they slot into the rest of the product assembly process).
Solving the “model boundary” problem
Using a traditional approach to modelling, determining models’ limits and knowing where boundaries lie is prerequisite to building a verifiable model.
Model boundaries are conditions — or set of conditions — which define the environment in which the model will be instantiated, and makes clear what variables, actors and factors are under consideration and which are not.
Take the example of consumer goods manufacturing. If you were to run a discrete event model covering how one of your production lines was performing (or going to perform) in terms of output, for example, it’s likely that at an extremely base level, the values you would choose to feature would include things like the cost and amount of input materials, the frequency and duration of any delays to effective passage along the line, and the standard of each unit as determined at the quality assurance station.
The resultant model, which would be likely to indicate how efficient or not your production line was, would definitely be helpful. It might indicate where you should reallocate certain staff members, for example, or at which stations you need to invest in new machinery or plant.
But what it fails to take into account is the effect that a whole other range of systems outside of the specific system in your warehouse, or even your company, is affecting each of the values you have incorporated into your discrete events model. The cost of your input materials, for example, is affected by a wide range of factors that are completely external to your sphere of control, like how choppy the market conditions are or how proximate to your warehouse they lie.
And to add an extra layer of model boundary complexity, the external systems affecting a business can then end up themselves being affected by management’s consequent decisions — which muddies the model boundary waters even further — creating positive and negative feedback loops which can be hard to capture in traditional simulations. For example: if our hypothetical consumer goods business buys overly-cheap raw materials for their kitchen equipment business, and the resulting manufactured goods break easily when they are finally used by the consumer, it would be reasonable to expect that dissatisfaction may result in increased eventual costs through returns under warranty, and lower lifetime customer values (with buyers switching suppliers for future purchases).
Agent-based models, however, allow for novel means of capturing ‘emergent’ phenomena, and modelling feedback loops in complex adaptive systems. Alone they may be overkill for modelling a simple FMCG manufacturing system, but a discrete event model embedded within an ABM, whose inputs and outputs at various stages can be influenced by the activity of the wider ABM model (universe), allows for combination of the two distinct modelling paradigms.
Best of both worlds: picking a base
Although at first glance the term “multi-method modelling” might suggest a system whose basic starting point takes elements of theory from more than one of the different modelling paradigms, it’s actually common for modellers taking the multi-method route to select one distinct methodology as their starting point and then integrate the components they need from the other methodologies as they go — rather than try to construct a hybrid model from the outset.
Deciding with what paradigm to begin may seem a little daunting, but one of the good things about taking a multi-method approach is that it gives you the chance to choose the paradigm which meets your needs best without locking you into that methodology for the whole model.
Through HASH, we hope to address the general lack of commercially available tools and standards suited for multi-method, multi-paradigm modelling.