Run a MonteCarlo sim...

[mlevison]

.. to plot the game actions and their affects. Vs the Gremlins and their affects. Which choices have the largest effect the most often.

This is clearly a special runner and its output will only be available to people who've paid for access to our learning hub.

@Xiphe how challenging is this? Both in dev effort and even run time?

[mlevison]

...eventually we could even make the game extensible - a learner wants to see what effect an action will have and they inject into the the game and see the results.

[mlevison]

@Xiphe Just how hard would the Monte Carlo simulation be to build?

[Xiphe]

I'm not familiar with Monte Carlo Simulations and just now read over the first parameters of the WIKI.

It would help if you could present me the problem to which "running a simulation" is the solution.

Simulations in general should be possible:

• The gameReducer is a pure function that is not tied to react. Therefore it could be invoked from any context and also in a NodeJS environment.
• A problem might be that the reducer itself has no validation if a given action is allowed at a given state of the game. (This is currently done by the UI/react part where later actions are just not displayed/enabled when they're not available)
• Speaking of extensibility, currently the actions and gremlins are hard-coded but it is definitely possible to untie them from the gameReducer and thereby allow running a simulation against a different game configuration.

Honestly: I have no idea how complex this is. Mostly because I have no experience with the simulation part.
Here are some (really really) rough estimates:

• Moving the game to state where simulations are easy to run: 1-2 days
• Creating a environment that can run such simulations: 1-2 weeks
• Allowing custom configurations for simulations: 1-4 days

[mlevison]

First up cool. Glad I could show you something crazy different. After that is what I do for a living.

To help - here is roughly what is in my brain. I would make one simplification to speed things along - each round the 3 actions are selected either from the current or previous rounds. From there Imagine running the game 20-30,000 times. All choices about choosing an action are made at random. That should be enough rounds to choose all of the choices many times over and see which results in the best outcomes the most often, since each set of choices will also be pitted against most of the gremlins.

A problem might be that the reducer itself has no validation if a given action is allowed at a given state of the game.

So it would seem that is the key improvement that I/we would need to make.

Creating a environment that can run such simulations: 1-2 weeks

Ok that's interesting - so I should only do this when the game is part of a bigger money making activity. Out of curiosity, what is hard here?

[Xiphe]

Cool! Your TL;DR explanation of the sim clicks a lot more than the wiki entry.

| Ok that's interesting - so I should only do this when the game is part of a bigger money making activity. Out of curiosity, what is hard here?

This is me trying to estimate in a field I have zero knowledge of. Can't tell what's hard here. Could also be doable in 1 day or 1 month.

But you tingled my interest now. I'd like to give this a shot just for fun.

How would you envision the UX? You said this should be s.th. that is "only available to people who've paid for access"

[mlevison]

Apologies I hadn't meant to call into question your expertise. My challenge is that I have a very rough idea of what this entails. I know it will be messy. My suspicion is this a little bit complex (how to handle output) and mostly complicated - how to hookup die rolls to our little game, how to store masses of data.

Basic cases for use:

• Me/You - did we build a game that is balanced - when you look over the logs do the actions you expect to win have the greatest effect

Paid use is more complicated. Some people will never need this and that's awesome. Imagine however a team that plays a few rounds and discovers they need more help. They want to understand which paths will likely lead to better outcomes and they want a hand getting there. They pay for access to deeper knowledge and ongoing coaching support.

This group won't get access to run their own MonteCarlo sims (I don't think). We might tolerate a runtime of 2-3 hrs, whereas they won't.

So less a UI and more of a command line tool that produces an output file or maybe just run in Observable. I suspect datasets will be huge and so we will need a database to store results.

We would record:

• Action selection by round
• Which gremlins happened
• Capacity and User Story Percentage, User Stories complete by round
• Final capacity, user story percentage and number of User Complete at game end

From that data you/me could sort/filter to find the actions that are consistently associated with better performance. That feeds into education.

In the short term access to the database would be through whatever tool we can work with - ex. Excel speaks to BigQuery or ....

In the long term I will find something that I can embed in a Learning Platform.
(BTW welcome to my truly mad brain - full of more ideas than one person can build in a lifetime).

[mlevison]

@Xiphe When it is practical could you start work on making the game playable in a headless fashion?

I gather looking at the code behind: https://observablehq.com/@danielfrey/monte-carlo-simulation-i will give us some idea. Possibly even better: https://medium.com/@quasimik/implementing-monte-carlo-tree-search-in-node-js-5f07595104df has done a better job implementing a montecarlo sim as a tree search.

I'm open to any solution that gets many many many game plays >1 million.

[Xiphe]

I'll be traveling back to Germany very soon. When things are a bit settled there I'd like to start with this. That might be in roughly 2-3 Weeks