Lecture 022

Reading

Boid-based simulation

• flaws:
  • assuming individual are simple to model
  • individual cannot calculate "most efficient action"
  • trade off between complexity and predictive accuracy (therefore the model aim to explain instead of predict)

Modeling

• individual -> macro

• macro -> individual

• macro outcome is more than sum of its part (analogy: metal harness when mixing But metalic molecules property depend on molecules)

• link micro and macro level explanation and phenomenon

• able to model chaos theory (but really? chaos is sensitive to initial condition)

• maybe can identify potential chaos system

Questions

1. macro-simulation can be variable-centric?

2. macro-simulation is variable-centric

3. micro, depending how complex the algorithm are

   • NN: still variable-centric because each layer lacks explanation what they actually mean
   • ABM: non-variable-centric because each pare meter has social meaning

4. can ABM better than individual psychology approach?

5. ABM can be verified by prediction of computer simulation whereas individual psychology can't be verified

6. Simpler model or complex model We should neither prefer a simpler, less realistic model nor a complex and realistic representation of agents. The assumption of a simpler model always results in a less realistic representation and a complex model always results in a more realistic representation of agents is unsound. Agent-based modeling is similar to Neural-Network modeling in terms of trying to adjust individual units (agent/artificial-neuron) to produce an overall effect by interactions of those units. In the field of Deep Learning, overfit refers to an inaccurate prediction resulted from adding the complexity of the model, and underfit refers to an inaccurate prediction resulted from having lower complexity than desired. The key to producing a good model, in ABM and Artificial Neural-Network is to find the sweet spot where a relatively low complexity model gives an accurate prediction. One might argue that prediction accuracy does not imply representation accuracy. However, I argue that representation should be guided (validated) by prediction if a relatively accurate prediction exists. In this case, a relatively accurate prediction exists because we can partially model the result produced by a chaos system by computer programs.

7. opposite arguments?

8. Macy and Willer suggest these models, due to their inaccuracy, should be used as explanation rather than a prediction.

9. Bruch and Atwell think the models can identify feedback loop.

10. "realistic represent": I don't see how they argue the model can realistically represent (just provide an additional way to verify complex phenomenon produced by simple rules)