In this post, we will look at the pros and cons of Agent-Based Models (ABM). ABMs are a common modeling tool use in computer simulations and can model some rather highly complex systems with little coding.
One strength of ABM is its ability to model heterogeneous populations. By heterogeneous we mean a sample in which the statistical properties of distribution are the same in different parts of the distributions. This is a key assumption in EBM and there are statistical tests that have to be done to assess that these assumptions are met.
ABM also allows for the discrete models rather than continuous models. Many actions within an ABM environment are “on or off” actions. This means that either they completely happened or they did not happen at all. EBM can have partial actions or outputs that would not make sense in the actual world. However, there are some ways around this through the use of a categorical dependent variable in an EBM. However, even in this situation, the EBM decision to pick on action over another is based on probability which can be less than 100%.
For ABM, another benefit is that the researcher does not need to have an understanding of the aggregate or big picture behavior of the phenomenon. These behaviors emerge from the rules that are developed by the researcher. With EBM, a strong understanding of the aggregate behavior or pattern is needed because statistics generally focus on aggregate patterns and not individual deviations from the aggregation. The exceptions to this my be unsupervised machine learning in which pattern discovery is the primary goal.
This leads to another point in that ABMs can incorporate randomness into the model since they do not know the patterns to expect. EBM tends to be deterministic, making it hard to account for randomness which is literally called error.
Lastly, ABM allows you to see a model develop over time because the data is generated during the simulation. Often, but not always, this is not possible with EBM as the data is already collected prior. this makes it hard to experiment with different scenarios
One of the strengths of ABMs is also their weakness which is that ABMs are not beneficial when dealing with homogeneous data. If the statistics in the distribution are all similar there is no need to identify individual agency or behavior. ABMs are focused on the differences in the individual and how these contribute to system patterns.
ABMs can also be computationally expensive. Each agent is moving about and demanding calculations from the CPU. When this number gets too large it can lead to a computer seriously slowing down. However, as technology improves this is less and less of a concern.
A final criticism of ABMs is there use of free parameters. Free parameters are essentially variables that the researcher must set to a specific value before running the model. In machine learning, these are called hyperparameters. Whether free parameters are a weakness or not is somewhat controversial since all models have certain things they assume in order to run.
This post examined the pros and cons of ABMs. In general, as one famous statistician said, “all models are wrong but some are useful.” This same principle applies to the use of ABMs.