Here are some of my favorite images from the 100×100 galleries – though I encourage you to check out the galleries and not take my word for it. For one, my opinion is heavily influenced by the fact that I have looked at countless thousands of images.
Unlike my previous (and rather outdated) Top 5, I’m not going to rank them and instead will show them in the order they appear in the galleries.
I’m a sucker for these levels where there’s a single, wavy focal point – if it gets the colors right, too. This one does, in a big way.
The yellow line outlines the main pattern in the center/right while the light blue line outlines a strong second pattern at the top. While they are both somewhat simple – there’s plenty more going on too.
A really complex pattern that has many different interesting parts while managing to avoid any really “junky” areas that are typical of levels like this. But maybe the coolest part about it doesn’t pop out immediately:
It’s not the first time that I’ve seen a bunch of dots lined up against a wall of steps – but in almost all cases the dots form lines that are mostly junk. These dots take a very different path to find their mates. The area is also extremely dark in the constellation map version – showing that the similarity filtering really likes it, too.
While many of my favorites (like the last two) are subtle – there isn’t much subtle about this one. It’s reminiscent of #2 in my original Top 5 – but far more interesting in my opinion.
There’s a short(ish) light blue line and a short(ish) light orange line outlining the “inner” central pattern, then moving outward – a single not so short light orange line that loops around 3 times and a single light blue line that does the same just outside that. There are of course other, darker lines in between which serve to highlight the craziness of these lines.
Another great example of a recurring theme of my favorites – tons of different interesting patterns and they all fit together almost seamlessly.
Ditto.
This is the one level that I included four images for in the set. That doesn’t mean it’s my favorite – it just means I had a very hard time choosing favorites for the level. The more I look at it though, the more I think that it just might be my favorite level.
A recurring theme in many of the levels in my galleries is a long set of steps emerging from the corner of the board that loop around a spiral (usually square) center. While the yellow line shows this almost perfectly – it’s no where near a perfect square, and the rest of the lines form this really funky, wavy pattern. The magenta version is equally good.
The more images I look at, the harder it becomes for a simple-looking pattern like this one to become a favorite. While I will admit the colors play a big role – the pattern itself is deceptively complex.
I used the cyan version of this as the thumbnail for the 3rd and last part of the 100×100 galleries. The cyan version does a better job of highlighting the main pattern in the center – but this one adds a significant counterpoint in the top left area that the cyan version mostly hides.
Notes
There’s an unintentional (and to me, slightly unfortunate) logjam of favorites at the end of Part 1 and Part 2 of the gallery – with 8 of the 11 images coming from the last 4 and 5 levels in each gallery. Part 3 doesn’t finish so strong in my opinion, though I am biased by the fact that the last 11 images came from levels that my program deemed less unique than the rest.
Constellation Maps
You can see another version of each of these images here: Constellation Maps. Below is an example.
How My Program Rates These Levels
When my program is finished generating levels and selecting the top X levels (usually 60, but for 100×100 it was 120 levels) – I randomize the order of the levels so that I don’t know how my program rated them. I won’t look at the ratings until after I’ve finished the gallery. Unfortunately, for the 60 additional levels I made images for after the initial set – I was not able to blind myself and was likely influenced by their relatively low score. I’ll ignore those levels (of which 1 of them made it to this page) and instead focus on the initial 120 levels.
I’ll separate the 120 levels into 3 categories:
- My favorites that are shown on this page (11 total)
- Other levels that made it into the gallery (31 total)
- Levels that didn’t make it into the gallery (78 total)
Top | In Gallery | Excluded | |
Total | 10 | 32 | 78 |
Avg Rank | 28.3 | 57.3 | 66 |
Geo Mean | 21.9 | 37.4 | 53.9 |
Best | 4 | 1 | 3 |
Worst | 72 | 120 | 118 |
Rank 1-10 | 2 | 4 | 4 |
Rank 11-30 | 4 | 5 | 11 |
Rank 31-50 | 3 | 6 | 11 |
Rank 51-100 | 1 | 12 | 37 |
Rank 101+ | 0 | 5 | 15 |
I set the rank ranges based on the stages in my similarity filtering:
- First, start with all the levels generated (20,109 in this case) and filter them down in stages to the top 200. Each iteration cuts the population in half. During this stage, I make some adjustments (especially in the earlier iterations) to increase the speed at the cost of some accuracy.
- Filter the 200 levels down to 100 levels and then 50 levels. The levels are split into groups of 25 compared to 10 in the first stage to improve accuracy. This particular stage also adds an additional “line uniqueness” factor to the equation.
- Compare all of the last 50 levels and select the top 30 from those.
The reason it finishes at 30 levels is that this is the typical number of levels for each board size within a level pack in the game. Even though I have been making images for at least the top 60 for a long time, I still let the program finish it’s course.
6 of the 10 top levels came from my programs top 30 levels and 8 of the 10 from the top 32 (31 and 32 are both here). Only 1 level was outside the top 50, and none outside the top 100.
Conversely, 52 of the 78 excluded levels came from outside the top 50. The one point where the excluded levels win: the 3rd ranked level didn’t make it into the set, while the highest ranked level on this page is only #4. The top two are both in the gallery, but didn’t make my favorites.
The below table shows the share of each category for the top 10, 30, 50 and 100:
Share Of | Top | Gallery | Excluded |
Top 10 | 20% | 40% | 40% |
Top 30 | 20% | 30% | 50% |
Top 50 | 18% | 30% | 52% |
Top 100 | 10% | 27% | 63% |
All | 8% | 27% | 65% |
The data almost perfectly matches what you would expect – the top levels have the highest share in the top 10 and top 30, with the share decreasing each stage after that, while the excluded levels start with 40% and increases each stage. The middle tier is more consistent not counting the 40% to 30% drop from top 10 to 30 (which equates to just 1 level).
I have to admit – given the sample size (especially including the top 10 category), the result is pretty lucky.
95×85
Gallery Here. I did the same analysis with the 95×85 gallery. The results aren’t nearly as perfect:
Top | In Gallery | Excluded | |
Total | 10 | 30 | 20 |
Avg Rank | 34.6 | 22.5 | 40.5 |
GeoMean | 30.4 | 15.2 | 38.0 |
Best | 8 | 1 | 13 |
Worst | 55 | 60 | 59 |
Rank 1-10 | 1 | 9 | 0 |
Rank 11-30 | 5 | 11 | 5 |
Rank 31-50 | 3 | 8 | 9 |
Rank 51-60 | 2 | 2 | 6 |
By share:
Top | Gallery | Excluded | |
Top 10 | 10% | 90% | 0% |
Top 30 | 20% | 67% | 17% |
Top 50 | 18% | 56% | 28% |
All | 18% | 50% | 33% |
Note that I haven’t published a top 10 for this gallery, but I did make a list of my favorites before looking at any of the rankings.
My favorites didn’t do well at all here, scoring less on average than 10 randomly selected levels. The levels excluded from the gallery fared significantly worse though – completely shut out of the top 12.
The middle category wins by a lot. However – it doesn’t take much to swing the balance. If I could trade just 2 of the top 10 levels from the middle to top category, the results would be very close to the 100×100 comparison.