← sources

My demo was too short and too easy. So I created a balance loop: bots, 100.000 runs a night, one change at a time

source: reddit (reddit:gamedev)published 2026-11-02 11:37ingested 2026-11-03 02:38

retained text

I’m getting my roguelite’s demo ready to release, and it has to be balanced first. This is the loop I use. (The video shows 9 bot runs at once)

1. My first playtest demo was short (3 acts) and easy: my best bot won 92% of its runs. So I made it 5 acts long, with more pegs and items that unlock as you level up, and for that I needed a process for balancing the game.
2. So I set up a cycle. On the first night, bots play the demo with no picture, just the simulation: about 10,000 runs on a normal night. The biggest night so far was 101,350 runs and 2.6 million drops.
3. Who plays, and how they think. One bot, two “brain” settings.
   1. The expert rehearses every throw. My physics is deterministic, so it can copy the whole game, try a throw in the copy and see exactly what would happen. It tries 24 random aim spots and throws at the best one. No player can do that, so it stands in for a very good player, not for a human. 
   2. The beginner tries only 3 spots, never saves money and knows nothing about the items. It’s close to random.

Buying works like this: for each peg on offer it tries the board’s most-hit spots and a few random ones, rehearses drops on each, and buys only if the board scores better than without it. The expert also saves for interest when it’s comfortably ahead, and takes items by what past nights measured them to be worth.

1. Every morning I get a page: the win rate for each bot, how many runs are still alive after each round, where they die, how far short a lost round fell, which twist rounds kill the most runs, and what every peg and item is worth. An item’s worth is measured by giving it at the start of a run and comparing against the same runs without it (same seeds, same luck), in rounds cleared. The runs in the video are reconstructions: the nightly has no picture, so I replay its runs in the game from the same seeds.
2. I sort every item into five piles: very strong, strong, normal, weak, very weak. A run is 15 rounds, and with the strongest item from the start, the bots got 3-4 rounds further on average than without it. The worst one actually lost them 1-2 rounds.
3. For each item outside “normal” I queue a few changes. I make it rarer (two items that won 38-40% of runs, against 6.6% without them, became rarer and pricier, because finding one should still feel great), buff it at two strengths, give a weak build an item that multiplies its own score, or cut it.
4. On the second night, the bots play each change on its own, against the same seeds. I know several changes together can give a different result; for now I keep it to one at a time, so each number means one thing.
5. The next morning, a new report shows what each change did, and I pick the ones to keep by hand. Some flags I leave on purpose: a rare drop thousands of times the target is the fun part.
6. Then the cycle repeats: another 10,000 runs measure the new game, and the loop starts over.

Of course, I’m testing the game myself in the process, and interestingly, I still enjoy it. I’ve heard that game developers stop enjoying their projects after a certain point because they know everything about them. But even now I find some cool interactions. The bots are great at telling two versions of the game apart; how hard it really feels is still my call.

I want to read 1-2 books on game testing and would really appreciate it if you could recommend any, especially if they cover my case. Thank you!

document fetched 2026-11-03 00:26 · canonical https://reddit.com/r/gamedev/comments/1xluvk0/my_demo_was_too_short_and_too_easy_so_i_created_a