Comments

[SarvagyaVaish]

Leave your comments here...

[xissy]

wow, this is amazing. inspired by your practical ML approach.

[iandanforth]

You should get a tapsterbot! https://github.com/hugs/tapsterbot

[Aaron1011]

This is incredible!

[joeyslater]

That's what's up man.

[dend]

Awesome job, Survy!

[halfdan]

Nice job - please add a proper reference to the source of the pseudo code though. It's clearly taken out of a publication.

[Giszmo]

I never did image analysis but I assume it to be trivial to do with a camera for your android bot. You said the image (screenshot) takes 2s to get to the PC? A cam should be much much faster. The image analysis would basically just need to scan the right side of the screen for green-notgreen-green. The timing is constant.

[ztl2004]

Dude, this is fantastic and it's what I be thought about for a long time, I ve noticed that u want to do this on mobile, I ve studied ios private Apis and I ve done the screen capture and touch simulation, do u think there is a possibility that we work it out

[bolte-17]

Any thought to adding either bird's current velocity (or as a proxy, time since last tap) to the state space? That seems to be the only missing parameter.

[ztl2004]

but I think it's hard to get

在 Feb 16, 2014，2:57 PM，bolte-17 notifications@github.com 写道：

Any thought to adding either bird's current velocity (or as a proxy, time since last tap) to the state space? That seems to be the only missing parameter.

—
Reply to this email directly or view it on GitHub.

[cbbayburt]

Actually, simulating the game's dynamics might lead to a simpler and more precise solution. the game doesn't really involve sophisticated decision steps which requires ML. Since it is really a pure physics problem, simpler solution depends on some simple observations though:

• Through observation, I found out that keeping the bird level requires a tapping period of 600ms.
• Let's say the bird's jump height is hb. So in original game, every tap makes the bird go up hb units, while every 600ms its height goes down hb units.
• Ascending and descending actions are achieved by simply modifying the tapping period (Smaller to ascend, lower to descend).
• Actual amount of ascend can be calculated as: dH = hb - (hb*ptap/600) From here, the required tapping period to achieve a specific ascend/descend amount 'dH' can be calculated as: ptap = 600 - (600 * dH / hb)

So the algorithm would be:

hb: The bird's jump height for a single tap, in other words, the amplitude of the bird's harmonic motion in level flight (it is constant and can be measured in means of pixels).
hBird: Height of the middle point of the bird's harmonic motion.
hObstacle: Height of the middle point of the space between the pipes.
ptap: Waiting period before the next tap.
dh: The height difference between the bird and the obstacle path.

for each immediate uncleared obstacle:
  while(obstacle_not_cleared)
    dh <- hObstacle - hBird
    ptap <- 600 - (600 * dh / hb)
    if ptap < 0 then ptap <- 0  //Gonna fall, tap immediately
    sleep(ptap)
    tap()

This algorithm can make the flappy lips fly forever. For android, instead of requesting .png screenshots which really takes about 1-2 seconds, you can analyze specific pixels in the raw frame buffer (some unix device file like /dev/graphics/fb0) which gives you enough speed to run the algorithm. But for that, you obviously need a rooted device.

[SarvagyaVaish]

Analyzing the "specific pixels in the raw frame buffer" is worth a shot! Thanks!
And I agree with the solution being nicer if I simulated the game dynamics, but I wanted to approach the problem using machine learning. Thanks for the solution though.

[savraj]

I'd love a deeper walkthrough of this -- maybe a youtube video.

[metaylor]

This is very cool. Good idea to pick a popular game and show that ML can solve it! I'm going to bring up your project as a discussion topic in the graduate reinforcement learning class I'm currently teaching.

http://www.eecs.wsu.edu/~taylorm/14_580/index.html

[SarvagyaVaish]

That is awesome!! I am honored. Thanks :)
May I ask how you found the link?