Stats
Levels Generated | 12006 |
Generated on AWS | 5266 |
Fast Settings (AWS) | 2093 |
Processing Time | 14 days |
Manual Time | 5 days |
Selected for Imaging | 60 |
In Gallery | 30 |
Images Generated | >3000 |
In Gallery | 55* |
*- the last 5 images were added later on after experimenting with a new color palette.
All of the levels for this gallery were generated with a new setting that adds some randomness at the cost of speed. This made it even slower than 100×100 – from about 1000 levels per day to 600 per day on my machine. That is why I experimented with Amazon Web Services. A virtual machine on AWS ran for 5 days, generating 5266 levels at a cost of 86 dollars. The rate of about 1050 levels per day is inflated since 40% of the levels were generated in 25% of the total processor time due to different settings that increased throughput.
Just before I terminated the VM, I changed 1 line of code that tripled the processing speed. At first, I didn’t think the change would have any impact on the levels – but I was wrong. The changed fixed a funny little bug that effectively (and unintentionally) added more randomness where I didn’t know there was any.
1858 of the 12000 levels were generated after the change, but only 3 out of 60 were chosen by my program (15% vs 5% of the total). This could have been random luck – but my program liked these levels even less in the earlier stages of filtering:
The first 3 passes reduced the total levels from 12006 to 1500 – with a 50% reduction each pass. Only 70 of the 1858 new levels made it this far – again from just over 15% of the total to about 5%. Oddly enough – according to my program, this makes the speed trade-off almost completely fair.
Of the 3 levels that did make it – 2 of them the computer ranked in the top 10 (albeit, at rank 9 and 10). The 3rd ranked at 47.
According to my own rankings – 1 of the levels is possibly my favorite, 1 of them is about right in the middle (it was the last image cut), and 1 of them didn’t even get a second look.
I suppose I will find out when the next set completes whether this code change is worth it or not.
Previous gallery: 95×85
Next gallery: 111×99