Structure

[hyunjimoon]

Hi all, I am trying to start with the following structure which may change dynamically (but will refrain from too frequent changes).

code_conti folder

:= static tools (.py, .r) and dynamic tales (.rmd or .ipynb). Classification for static tool follows Bayesian taxonomy from this very useful! paper from my SBC collaborator. I hope this discussion can be developed in another discussion thread I started here.

1. PAD: empirical analysis e.g. Vensim-Stan inference builder `draws2data()`, `data2draw()`, `sbc()`
2. PD: theoretical analysis e.g. structural dominance analysis assuming asymptotic
3. Data4DM: data for dynamic model and decision maker centered model (SD2nonSD), policy parameter prescription e.g. SOPS, BATS
4. DM4Data: dynamic model and decision maker for data-driven model (nonSD2SD), policy parameter prescription e.g. temporal GAN, causal identification
5. Auto: unsupervised

Shall I rename the 3 and 4 to Stats2Sci, Sci2Stats?
Detailed explanation on PAD, PD below.

wiki_week folder

:= seven documents each explaining model combination of P, A, D in system dynamics concept

1. Branching Bayesian: Bayesian taxonomy
2. Merging to Bayequentist: Statistical equivalence of Bayesian and frequentist and their role in workflow
3. Bayequentist Dynamics with SBC: implementation of SBC in dynamics and cases studies on SIR, Population, supplychain dynamics 
4. System Structures: eight representative P models
6. Merging Algorithms: Tracing the history and trend of A models
7. Branching Behavior: Bayesian inversion of observed to unobserved parameters (switching cause and effect)
8. Simulated-data experiment

Tuning and classification discussed in discussion #5 should be documented in 6 or 7.

model_month folder

:= twelve classification of models, depending on

• material/energy/information open-closed
• purpose of the system (innovative vs regulative)
  added to eight representative models covered in 879 class (1~7).

1. mng(diffusion)
2. mng(project)
3. mng(capacity)
4. mng(market)
5. mng(material)
6. mng(service)
9. mng(city_world)
10. mng(energy)
11. mng(information)
12. mng(chain)
13. mng(variation)
14. mng(game)

Models from BD and AMD textbook classification to the above in progress.

seminar_season folder

:= relevant seminar summary or lecture notes that I took or taught. Especially the summary contents I upload relevant to the class I am TAing for Hazhir will be useful.

---

Excerpts from the paper on classification:

1. PD models := combination of P model and observed data $\tilde{y}$

• $(p(y, θ), \tilde{y})$
• represent the ideal endpoint of Bayesian inference. However, because we can rarely compute the marginal likelihood p(˜y) analytically, we do not have access to the actual PD model outside of textbook

2. PA models := combination of a P model with a posterior approximator A

• $(p(y, \theta), p_{\mathrm{A}}(\theta \mid y))$
• useful for confirming computational faithfulness of a workflow, for instance, via SBC or assessing adequacy of a model for answering a particular research goal [231]. Importantly, type of P model (i.e., $P_E$or $P_I$) will determine or necessitate the choice of a particular approximator A.

3. PAD models := combination of P model, a posterior approximator A, and observed data D

• $(p(y, \theta), p_{\mathrm{A}}(\theta \mid \tilde{y}), \tilde{y})$
• aim to approximate the corresponding PD model by means of a suitable approximator A, whereas the amount of data D, together with the type of P model, will largely determine the choice of approximator. Misspecified models might still cause considerable difficulties and convergence problems for MCMC methods in practice. Thus, any trustworthy approximator should be equipped with diagnostics signaling improper convergence or invalid inference queries