Map for vensim-stan connection

[hyunjimoon]

All codes including demo on prey-predator and inventory management will be actively updated in https://github.com/Data4DM/stanify. The gap between system dynamics and Bayesian statistics cannot be filled in one semester, but we can increase its filling rate together! Please leave error report or question in stanify repository as issue here. I am preparing for demo for 15.879 class, upcoming Friday. Note that this post is not official and I am responsible for all potential wrong facts.

I summarized below lessons I learned during three months of vensim - stan connection. I strongly encourage you to leave comments which will help me allocate my resource on the features that can benefit you the most.

Variable types and end goal

Customized computation model

• The Pinocchio Principle: A model that is created solely for computational reasons can take on a life of its own
• James' comment of potential hallucination of sine wave as constant function with downsampling (large savepars)
• Tom's comment on sine wave vs exponential growth in #4. Tuning window, the number of time-step, for behavior classifier (or pattern recognition) #11

Relative, not absolute time_step

• originally, time_step was not translated to stan file; stocked_flow structure is how time_step first got introduced to stan

• new parameterization via fraction of time step over correlation time 0 < tolerance = time_step/correlation_time < 1 (just as adjustment time has the same information as fraction). Time step is not translated to stan model. As time_step is an artifact whereas correlation_time is physical, modelers should customize tolerance according to their needs. At one side of the tolerance spectrum, miniscule tolerance will make your model impractical from heavy computation (10 simulations per sec.) whereas large tolerance (i.e. larger time_step) will make your model unrealistic.

• helpful explanation on tolerance: "two-sided tolerancing is like goal posts in a football game; all data within tolerances are equally acceptable"

Data as stock

• For the highest accuracy, the format of data should be stock which is natural considering its 'observed, measured, reported' character
• this can formalize imputation framework for missing data (if data <> NA, 0, distance)

Stock vector

• For the highest accuracy, the initial value should be estimated parameter.
• stock_init value is HIGHLY implementation-sensitive e.g. affected by order of computation.
• stock _vector is affected by tolerance e.g. downsampled target simulated data, vector to be matched
• with unrealistic equilibrium assumption, specify reference_throughput parameter, then let all stocks initialized around parameter (Tom's rec). Focus on variation design after initializing. e.g.:

reference_throughput = 10 ~ widgets/week
customer_order_rate = reference_throughput + customer_order_variation ~ widgets/week
customer_order_variation = (random or test input) ~ widgets/week
wip_inventory = INTEG( production_start_rate - production_rate, (reference_throughput)* manufacturing_cycle_time + safety_stock) ~ widgets
safety_stock =  reference_throughput * (safety_stock_coverage + order_processing_time)  ~ widgets

Need for convergence diagnostics

(upcoming)

---

my questions
Q1. Can stock_init be affected by tolerance?
Q2. How to select tolerance, a sweet spot for both practical and realistic portrayal?

@jandraor after your research summary tomorrow I wish to discuss the above (if time permits).
@tomfid the above is the summary of our discussion last week; may I ask for some feedback?

[hyunjimoon]

Other agendas with Jair and Paul

• Should we keep two separate stan function files for models with different numeric assumtion? #64
• symmetric vs asymmetric generator and estimator Data4DM/stanify@887352b#diff-663a190e1d5be1099a028bac38c1242919ca53f72d54e543a68f2df6cee84fa2

[hyunjimoon]

@tomfid I think it's time for our first cookbook! I will share this map of Bayes-SD to will help students stress-test stanify this monday. If it isn't too much to ask could you help me by filling in the vensim column?

For the collaboration, I put the code for table here.

Our previous discussion on Nathaniel's Rosetta stone ml - biostats map might be second-order relevant. #23 (comment)