Show HN: Unity 3d Neuroevolution Flappy Bird | Hacker News Reader

Show HN: Unity 3d Neuroevolution Flappy Bird

github.com

68 pointsgalleanv16 comments

Discussion of the original: https://news.ycombinator.com/item?id=12917538

Seems to have the same problem as the original where its not actually converging towards a solution but randomly stumbling onto it.

Is there a qualitative difference? I was under the impression that neuro networks worked by converging to a solution BY randomly stumbling into it? Is my rudimentary understanding off this off?

Actually, I think that the one that you linked to does converge towards a solution. This new one does not.

This is pretty awesome. I am curious though, does this in any way use the same algorithm as the project that inspired this (https://xviniette.github.io/FlappyLearning/). That one appears to be vastly quicker at learning (reached over 13 million and counting at generation 18).

I think that environment in Unity version is more complex, it uses "real" physics (box2d) to simulate the world. JS version use very simple collision detection. But I am not 100% sure (this is my beginning in AI world)

It made it to 25k by gen 160. Then it deviated an started losing again. Gen 250 was virtually unstoppable, hitting 120k. Watching gen 251 right now, and it just crushed 100k... 150k... 200k... 242k. Wow. And Gen 252 is killing it too. Seems to have found a sweet spot.

The Asteroids demo is cool too:

https://xviniette.github.io/AsteroidsLearning/

AI Lab at UTexas Austin has more Atari 2600 NEAT demos as well. The Kung Fu Master "button masher" style makes me think a Street Fighter V agent would be hilarious to see in Human vs AI competitive action!

http://www.cs.utexas.edu/~ai-lab/?atari

For a proof of concept, this shows the capability of ML in a web browser. Creating an ad hoc cluster of a hundred webrtc peers would yield a formidable (but slow) meta learner. And just to put it in perspective: sentient.ai's petascale darkcycle cluster consists of ~1M CPUs and ~1k GPUs:

https://www.sentient.ai/platform/scale/

Distributed computation seems to give natural evolution a competitive advantage: it is by design already parallel.

Reference for further study: Evolving Deep Neural Networks

https://arxiv.org/pdf/1703.00548.pdf

Why is there only ever one (sometimes two/three) decent players every generation? Most will end up dying after one or two poles... how come once the algorithm gets decent there are many more making through the game along with the best one?

(Also, I'd like "number of taps per game segment" be a weight... fall until the very last second then a couple of quick taps - none of this "spamming to keep level" stuff: where's the style? ;)

I am beginner in all of this and also I would like to know why this is happening. In this example every generation have 20 new genomes (birds), 3 best stay from previous generation (elitism), rest are effect of mutation. Maybe mutation rate/range have to be tweaked. Current solution is best what I found so far.

It seems like half fall below and/or above the screen limits every round. I may be incorrect but you would think evolution after a round or two would take care of that type of issue.

After being stuck at ~24k for a long time on the 1051st round it scored 83170. Seems to learn very slowly, perhaps the new generations are too random?

This is fun! Does it learn the same from many short, low scoring runs as it does with fewer, longer, high scoring runs?

It it example it learns form best 3 runs

It seems to get worse if run at 5X or 10X speeds.

It is because physic simulation in unity gets worst when time speeds up

my experience is at 50 the first smart bird appeared, then at 121 they became consistently good.

Reply on news.ycombinator.com