• Deep learning made Artificial Intelligence pop (again)


• Data Science, Big Data and low-cost/high-performance computing


• Simulation associated to "old" Artificial Intelligence"


• Modelling is a fitness contest? (pun intended)


• When to use: purposes and context


• Simulation in archaeology:
  "Can you validate your model?"*
  

  Epstein, J. M. (2008). Why Model? Journal of Artificial Societies and Social Simulation, 11(4), 12.

how similar or different are Machine Learning and simulation models?

the
Epistemological context

Diagrams available at https://github.com/Andros-Spica/modelling-simulation-graphs
Icon assets from "The Noun Project" (thenounproject.com)

how similar or different are Machine Learning and simulation models?

the
Epistemological context

Diagrams available at https://github.com/Andros-Spica/modelling-simulation-graphs
Icon assets from "The Noun Project" (thenounproject.com)

Simulation for
the creation of training datasets

Jareño, S. J. N., Helden, D. P. van, Mirkes, E. M., Tyukin, I. Y., & Allison, P. M. (2021). Learning from Scarce Information: Using Synthetic Data to Classify Roman Fine Ware Pottery. Entropy, 23(9). 10.3390/e23091140

• 5373 images of 162 terra sigilata from Roman Britain
• A classification problem
• Simulation (or procedural generation) of 1000+ images per class using Matplotlib and Blender (experimentation)
• Taking advantage of deep learning, while minimising curse of dimensionality

Simulation datasets are (potentially) "Big Data"

Machine Learning for
model generation

• Immense (and growing) collection of Large Language Models (LLMs) of various kinds and domains
  e.g. ChatGPT, Google Bard, GitHub Copilot, etc.
• Conversational technique as your own "think thank"
• Variable fluency with programming languages
  e.g. Python > NetLogo
• Potential of using Web and local files data
• Technologies in "beta" - e.g. AgentGPT
• Iterative process still requiring a lot of know-how both on simulation and ML

You are not alone anymore!

Machine Learning for
pattern detection in simulation output

• Less new and hype, yet more important
• Realm of exploratory and inferential statistics
  e.g. linear models, multidimensional scaling (PCA, CA, etc), clustering (hierarchical, k-means, etc), Bayesian statistics, etc.
• Good, fairly consolidated practice:
  sensitivity analysis with Random Forest (or similar) on randomised exploration of parameters
• Many sources available (publications, libraries/packages, scripts)
• Always remember the nature of simulation output (e.g. account for model and experiment biases)

After 10,000 runs...
the headache comes

Machine Learning for
parameter calibration and optimisation

• Old problem, but still presenting challenges
• Effective parameter calibration is one of the missing link between "toy model" and "digital twin"
• Several quite consolidated algorithms for linear and non-linear models
  e.g., those available with optim() in R, Genetic Algorithms in OpenMOLE, etc.
• Challenges come with number of parameters and stochasticity

Threading an invisible, multidimensional landscape

Machine Learning for
model selection + pattern detection in output + parameter calibration and optimisation

Carrignon, S., Brughmans, T., & Romanowska, I. (2020). Tableware trade in the Roman East: Exploring cultural and economic transmission with agent-based modelling and approximate Bayesian computation. PLOS ONE, 15(11), e0240414. 10.1371/journal.pone.0240414

• Trade of tableware from eastern Mediterranean (Hellenistic and Roman periods)
• 8730 fragments, 5 types, 178 sites (presence/absence per type-site)
• A cultural evolution problem
• Simulation of cultural transmission algorithms (three hypotheses) coupled with a market economy model, producing spatial distributions of cultural traits of traders of each settlement
• Stochastic exploration of parameter space for each algorithm, evaluated in light of empirical data, through Approximate Bayesian Computation (ABC) + Population Monte Carlo (ABCPMC)

Cane turns into cannon!

Machine Learning for
surrogate model generation

• Input-output relationship in mechanistic models is approximated with non-mechanistic (low fidelity) and computationally cheaper models (e.g. polynomials) (Forrester et al. 2008)
• Higher simulation speed and lower computational costs necessary for:
  • more extensive parameter/scenario exploration
  • multi-objective calibration and optimisation
  • increased model scale and complexity
  • simulation using real-time data (e.g. digital twins)
• Machine learning models can:
  • build surrogate models given pre-defined building blocks (e.g. Greig & Arranz 2021)
  • serve as surrogates themselves, i.e. trained to predict simulation runs with limited amount of simulation data (e.g. Angione et al. 2022; Pfrommer et al. 2018)
• Caveat/pitfall: justified when there is high confidence on mechanism underlying model behaviour

KIDS meets KISS

Machine Learning for
simulation model components

• Not strictly new in distributed simulation methods
• Further exploring ABM roots and use in robotics
  (Multi-Agent Systems)
• Many learning algorithms directly compatible with ABM, e.g.:
  • reinforced learning (e.g. Angourakis et al. 2015)
  • evolutionary algorithms
    (e.g. Sarker & Ray 2010, Li et al. 2023)
  • neural networks!
• Several ABM frameworks are compatible with ML
  e.g. Belief-Desire-Intention agent architectures, OpenAI experiments with emergent agent behaviour
• Many topics friendly to both simulation and ML components
  e.g. collective action, cultural evolution, markets, etc.
• Caveat/pitfall: optimal behaviour might not coincide with the historical/target behaviour

Simulation (artificial) intelligence

• More domains, true transdisciplinary (design)
• Faster workflow (implementation, debugging, optimisation)
• Higher computability (size)
• Better calibration (input data)
• Better validation tests (output data)
• More expressive results
  (analysis & visualisation)

⇛ Tractability of complex processes

Caveat: higher computational costs, expertise requirements, hybrid teams desirable

Angourakis et al. 2022, Quaternary | model repository: https://github.com/Andros-Spica/indus-village-model

⇑ simplification of Land Water model (runoff calculation)
(surrogate model generation)

use of ML component in agent decision-making ⇒
(model component)
sensitivity analysis per each submodel ⇒
(pattern detection)

⇐ validation of Weather and Land model to known present conditions
(parameter calibration and optimisation)
⇐ Conversion of paleoenvironmental and paleodemographic data (proxies) to model (mechanistic) variables
(pre-processing/selecting input data) ⇓