support for binomial active/dormant in storage

[nuclearkatie]

Adds the ability to use a binomial distribution to active and dormant cycles which requires one parameter, the probability that any individual time step would end the current cycle. This is specifically useful for disruption modeling, where you can assume that at every time step there is some very small chance that the agent goes offline (i.e. dormant) and vice versa for dormant.

This is a negative binomial distribution (since we're sampling the number of trials until n successes, rather than sampling the successes in n trials), in the specific case where the number of successes is always one. This is equal to a geometric distribution, but since negative binomial is more generic that's what is implemented in cyclus#1841 and used here. Note that like a coin toss, every time step is independent and has the same probability, active_buying_end_probability (or dormant)

Requires cyclus#1841 to be merged first

[github-actions]

Build Status Report - dfd4e23 - 2024-10-31 02:27:52 +0000

Build FROM cyclus_20.04_apt/cyclus:latest

• Cycamore: Success ✅
• Cymetric: Success ✅

Build FROM cyclus_20.04_apt/cyclus:stable

• Cycamore: Failure ❌
• Cymetric: Skipped due to upstream failure ⚠️

Build FROM cyclus_20.04_conda/cyclus:latest

• Cycamore: Success ✅
• Cymetric: Success ✅

Build FROM cyclus_20.04_conda/cyclus:stable

• Cycamore: Failure ❌
• Cymetric: Skipped due to upstream failure ⚠️

Build FROM cyclus_22.04_apt/cyclus:latest

• Cycamore: Success ✅
• Cymetric: Success ✅

Build FROM cyclus_22.04_apt/cyclus:stable

• Cycamore: Failure ❌
• Cymetric: Skipped due to upstream failure ⚠️

Build FROM cyclus_22.04_conda/cyclus:latest

• Cycamore: Success ✅
• Cymetric: Success ✅

Build FROM cyclus_22.04_conda/cyclus:stable

• Cycamore: Failure ❌
• Cymetric: Skipped due to upstream failure ⚠️

Build FROM cyclus_24.04_apt/cyclus:latest

• Cycamore: Success ✅
• Cymetric: Success ✅

Build FROM cyclus_24.04_apt/cyclus:stable

• Cycamore: Failure ❌
• Cymetric: Skipped due to upstream failure ⚠️

Build FROM cyclus_24.04_conda/cyclus:latest

• Cycamore: Success ✅
• Cymetric: Success ✅

Build FROM cyclus_24.04_conda/cyclus:stable

• Cycamore: Failure ❌
• Cymetric: Skipped due to upstream failure ⚠️

[gonuke]

I think this looks OK for the example you have, but not sure this is the best way to accomplish the disruption modeling scenario. I was thinking more about something that had a regular active/dormant cycle, but could be disrupted for some different amount of (longer) dormant time with some probability.

[nuclearkatie]

Ah, interesting. So that's kind of like a negative binomial on top of another distribution?
Sample from typical distribution for active/dormant length, then sample from a regular binomial with trials = sampled length to see if it is disrupted? Or am I still misunderstanding the logic you're interested in having?

[gonuke]

I think I'm thinking of something simpler. Sample from a Bernoulli (binomial with N=1) distribution with a high probability (p = 0.99), and use a nominal dormant length (e.g. 10 days) when successful (= 1 success, happens most of the time) and a disrupted dormant length (e.g. 150 days) when non successful (=0 success, happens rarely).

Of course, this could just be done by sampling a random number and comparing to p, but maybe it makes more sense in your Boost distribution framework.

[nuclearkatie]

Ok I started working on the plane but I want to check in one more time before getting a working version. Basically, this would require a new understanding of dormant_dist in the buy policy. Right now, the dormant dist is always sampled directly as the dormant period length. For a disruption like this, you're describing a dormant dist that's used as as a binary, and then another parameter (or dist) would actually be used to set the next_dormant_end. I could do this by having a new buy pol Init with a new var (actually dist because at that point a dist is the same amount of work as a var and storage could just used permanently fixed dist for your notional length anyway), but then you need more special handling similar to cumulative_cap because now the dormant dist is not just something that can be sampled... Unless I make a wrapper dist? Either way, more handling is needed in Cyclus and not just Cycamore at this point to implement your suggestion

[gonuke]

It does need something new in Cyclus and I would implement it something like this: cyclus/cyclus#1827

Could also use a binomial distribution under the hood

[gonuke]

It does need something new in Cyclus and I would implement it something like this: cyclus/cyclus#1827

Could also use a binomial distribution under the hood

Of course - this is not building properly - but I use it to show the concept.

[gonuke]

It does need something new in Cyclus and I would implement it something like this: cyclus/cyclus#1827
Could also use a binomial distribution under the hood

Of course - this is not building properly - but I use it to show the concept.

It builds now.... I think I'm using the binomial distribution correctly. I expect it to return either 0 or 1, and it should return 1 with p_success.

[nuclearkatie]

Ok, I added FixedWithDisruption as an option to both active and dormant cycles using your new distribution in cyclus#1827 (obviously requires that to be merged first). Hopefully I've made the docs clear enough so people can decide between the two skewed cycles

[nuclearkatie]

Passing on latest now

[gonuke]

I think there is one optional parameter missing in the comments. You also didn't include the list of distribution types in the top of the header - I think you mentioned that today.

Finally - one optional request for formatting in comments.