Stop the side-scroller hero from rescuing the princess! The "player" is an AI that uses neural networks trained with the Neuro-Evolution of Augmenting Toplogies (NEAT) algorithm: http://en.sfml-dev.org/forums/index.php?topic=9897.0
You need to have the VS2010 redistributable package installed to run the game.
Select the items you want to use to prevent the player from making it to the princess, place them appropriately, and then hit "play" so that the AI can try to beat the level. If it finds a solution before it reaches 100%, you lose!
EDIT: I had to perform a post-compo bug fix, since the screen just turned black for some people, making it unplayable. It should work now.
Fun game. The AI sometimes got stuck in a tile but that was when it was UTTERLY DEFEATED anyway. BWUHAHA!! Simple but correct sound effects. Cool use of AI as a protagonist.
Awesome idea. As I understand I can only complete 7th level by cheating.
C
CireNeikual18 Dec 2012, 16:52
@tumoxep: Thanks for playing! How were you able to cheat? The last level can be done with only one item, since I made the spikes a bit too wide accidentally. But, it can also be beaten the way it was intended :)
Very cool use of NNs. I'll check out the paper at some point, but a comment/blog post about how you implemented this in the game would be great - what type of NN did you use and how was it structured? What did you use for inputs/outputs? How did you decide a sensible direction to move the weights, rather than "you died, try something else"?
The game itself is pretty fun, although the other aspects are nothing special - it's the technology that stands out here.
C
CireNeikual18 Dec 2012, 18:18
@Sim: The neural networks have 27 inputs. 2 of them are the x and y distances from the goal (signed), and the remaining 25 form a 5 by 5 grid around the bot that tell it what the tile types are (sort of like vision). There are 3 outputs: move left, move right, and jump. They are all boolean (true if neuron output > 0.5, false otherwise). The networks are recurrent. The structure is evolved as well, so it varies depending on the task. The weights are perturbed by a random amount in a random direction. The bots are evaluated based on how close they got to the princess. Death is not a cost. The version I uploaded ran for only 30 generations, although I should have probably made it go for longer. I was at first concerned about performance, but when I took it out of the IDE it ended up being pretty fast anyways. I chose a average species populations size of 40.
Fantastic AI, I'm always impressed when someone can make good AI, it's really difficult. Thanks to this idea can work and... Well, just a really good game!
Noooo...Princes was not saved :( Well played CireNeikual! Nice puzzle concept, I had fun. Graphics could use some tweaks - less visible grid and some tile variations, but that's just a minor complain. Good game, I would gladly play longer version of it.
I really like this! Neural networks are my favorite model for machine learning applications. :P Very well implemented into this theme-appropriate game. Great work!
Awesome use of AI! Some of the levels were a little too obvious, but it was pretty cool to see how the AI handled situations. It might be cool to have more blocks and see how the AI handles perhaps 6 different obstacle types.
This is an original gametype! And its fun as well. Its a shame that when you have the right solution right away, you don't see the AI trying all these different tactics in vain. But I understand that just watching the AI for hours on end wouldn't be very fun.
I ended up finding a one tile solution to the final level too, in a way that seems to have tricked the AI, but would otherwise be really easy to solve. SPOILER HERE: http://imgur.com/6MNhy
C
CireNeikual22 Dec 2012, 12:42
@ethankennerly That's really weird, I thought I fixed this problem. I will have to look into it further I guess.
@Norgg Yup, I found this one too after I submitted the game! The hit box for the spikes are a bit too large, and the AI has no sense of velocity, leading it to die there. Both of these things are fixed in the post-compo version though, check it out!
This was great! This idea only works with a good AI, and yours was quite adequate. The only level that I found to be a puzzle was #5, but every one was fun. Gameplay wise, this is one of my favorite games of the compo so far.
Comments54
> Dec 18, 2012 @ 4:35am
> when i run it all i get is a solid black window :(
That happened to me too. I run a Win8-x64 system.
As I understand I can only complete 7th level by cheating.
The game itself is pretty fun, although the other aspects are nothing special - it's the technology that stands out here.
It is so simple, but the AI make it so complicated and you can make so many puzzles with it...
More levels please!
You could have named your zip/exe same as the entry, I got troubles finding it again through LD search.
I'll try again on a different computer when I can.
A Puzzle game with cool AI!
It's fun to play, and I am still trying to pass the level 5 lol.
Good job!
Its a shame that when you have the right solution right away, you don't see the AI trying all these different tactics in vain. But I understand that just watching the AI for hours on end wouldn't be very fun.
I ended up finding a one tile solution to the final level too, in a way that seems to have tricked the AI, but would otherwise be really easy to solve. SPOILER HERE: http://imgur.com/6MNhy
@Norgg Yup, I found this one too after I submitted the game! The hit box for the spikes are a bit too large, and the AI has no sense of velocity, leading it to die there. Both of these things are fixed in the post-compo version though, check it out!
On the negative side, the sound and graphics weren't very good, but the excellent AI makes up for it. I would've loved more levels too.
Overall, great job!