Read on Twitter
Sensitivity Analysis is the latest addition to the toolset, providing broad & granular insights from the solver
In this thread, I'll go through:
- What is Sensitivity Analysis?
- Why & Where it holds value
- Explanation of Noise Injection
- Special Notes https://t.co/UZdCrqo9ujfplreview.com
In this thread, I'll go through:
- What is Sensitivity Analysis?
- Why & Where it holds value
- Explanation of Noise Injection
- Special Notes https://t.co/UZdCrqo9ujfplreview.com
What is Sensitivity Analysis?
- Runs 5-1,000 solves using projections & user settings
- Adds "coherent" noise for each solve
- Outputs top selections across all solves
Small margins can greatly affect the single top solve and skew overall data interpretation
- Runs 5-1,000 solves using projections & user settings
- Adds "coherent" noise for each solve
- Outputs top selections across all solves
Small margins can greatly affect the single top solve and skew overall data interpretation
Why use Sensitivity Analysis 1
A single solve can sometimes miss the forest for the trees, especially in GW1 where many practically equally strong squads exist. The top solution could involve/exclude players by slim margins, who'd otherwise be common choices
A single solve can sometimes miss the forest for the trees, especially in GW1 where many practically equally strong squads exist. The top solution could involve/exclude players by slim margins, who'd otherwise be common choices
Why use Sensitivity Analysis 2
It offers a sense of scale too. For example, Salah might not emerge in noiseless top solves – leading to assumptions. Yet, ~30% noise-injected sims feature him, he can still be in the conversation
Sensitivity Analysis gives us greater context!
It offers a sense of scale too. For example, Salah might not emerge in noiseless top solves – leading to assumptions. Yet, ~30% noise-injected sims feature him, he can still be in the conversation
Sensitivity Analysis gives us greater context!
"Coherent" Noise Injection 1
In reality beliefs don't need perfect alignment with models; noise injection allows for this coexistence.
Random noise is added to team strengths, player stats, xMins, solve settings, etc. coherently, embodying diverse world views
In reality beliefs don't need perfect alignment with models; noise injection allows for this coexistence.
Random noise is added to team strengths, player stats, xMins, solve settings, etc. coherently, embodying diverse world views
"Coherent" Noise Injection 2
Coherent" means maintaining realism. For instance, a teams total xMins stays calibrated; players don't surpass 95 xMins etc.
Team strength changes also affect entire defences for CS%, etc.
Noise injects while staying realistic, not inducing chaos
Coherent" means maintaining realism. For instance, a teams total xMins stays calibrated; players don't surpass 95 xMins etc.
Team strength changes also affect entire defences for CS%, etc.
Noise injects while staying realistic, not inducing chaos
Special Note 1 (xMins Noise)
There is more uncertainty for "medium" xMin players – the ouputs from sims may emphasize potential with higher xMins belief (ie. JP picked in instances of higher belief)
You can disable/reduce xMin noise if preferred (as in image)
There is more uncertainty for "medium" xMin players – the ouputs from sims may emphasize potential with higher xMins belief (ie. JP picked in instances of higher belief)
You can disable/reduce xMin noise if preferred (as in image)
Special Note 2 (Evaluations Input)
The evaluation input is only relevant to non-WC/FH runs
Effectively you run the 5-1,000 simulations and the most common top solves are found- following this you can calculate the evaluation score of the top "N" options in the noiseless case
The evaluation input is only relevant to non-WC/FH runs
Effectively you run the 5-1,000 simulations and the most common top solves are found- following this you can calculate the evaluation score of the top "N" options in the noiseless case
Special Note 3 (Speed Boost)
For (slightly) faster results, set the number of simulations/evaluations as a multiple of how many threads your CPU can handle. ie. 25 sims will take ~33% longer than 24 sims for a CPU that can handle 8 threads
For (slightly) faster results, set the number of simulations/evaluations as a multiple of how many threads your CPU can handle. ie. 25 sims will take ~33% longer than 24 sims for a CPU that can handle 8 threads
I hope this thread gives some clarity on what exactly the sensitivity analysis on the site is and what kind of information and value you can get from it- I'm certainly finding it useful for gaining broad perspectives for GW1
• • •
Missing some Tweet in this thread? You can try to
force a refresh
