Heading back home after a month of bathroom building be like.
100 solutions tested in 4 weeks. Worst. Best. Difference between worst and best: 0.18 m/s outlet velocity vs 3.2. This is the third iteration, so the top 10 solutions from the previous iterations are seeded into the next iteration. The first iteration was 750 low resolution simulations to hone the general features (particle size: 3 mm). The second iteration was 200 medium resolution (2mm). This was 100 at high resolution (1mm). 235.2 kWh at 12 cents per kWh. 28.2 USD spent. Best employee ever. Bonus, best velocity per rotation.
A gradient descent algorithm like this could find the ideal parameters for a balanced game in no time. Don’t even tell players that it’s running. Just launch it and let it tune parameters until GM is balanced. It will automatically hone in on the changes needed for balance utopia by generating balance patches, assigning 100 random games to a patch, and measuring the correlation to winrate. It would do this for 100 balance patches and in total have 10000 games that define how each parameter impacts balance. Then it generates a balance patch that based on that data has the maximum probability of equalizing GM. It assigns 100 more games and if confirmed it’s pushed to the whole ladder. This happens automatically, without bias, and is dirt cheap to run. It takes zero effort from any design team or shadow-counsel of protoss loving baboons.