A non-DQN implementation playing flappy bird...learns fast enough for any cpu!
Fast, more human like style learning experiment. For people who don't have a titan X and want to train game playing bots!
So we know we can make DQN systems play games.
problem is, they are slow to train (10 hours on a titan x... i don't have a titan X!)
So is there a way we can make it learn as fast as a human? (took me couple of hours to play it a little, oke, still not good at it)
How do we do it?
- we read instructions, than make pictures in our head of what can go wrong(or is that just me?
- try to avoid that
So, if we can make a neural net that can translate written instructions to pictures of what can go wrong and train from their, we
should get fast results. (and have H-learning, human learning!)
This is the seccond part of the solution. Still working on a crash image generator that takes a screenshot and generates all that can go wrong
So can we train a bot to play flappy bird just from images of flappy bird crashing?
Turns out it is not only possible but it learns very fast. One hour on an I5 CPU should be enough to get some decent results
After about 35 times (10 games per try) it is already starting to get the hang of it! (after 50 it looks nice)
So that is about 15.000 frames of training to get some results (not perfect yet at that stage) compared to far over a million on DQN networks.
Try it...forget about your graphics card, your cpu will do just fine!
(just create the anaconda env with conda-env (provided) and run MLPFlap.py)
No working on the second part...taken an image of the background and of the bird (that is normally in human documentation) can i generate the crash pictures?
If so we are getting close to learn a machine play flappy bird the way you learned it!
it depents on pygame, keras, pandas, sci-image. The ananconda enviroment is included as conda_env just import that and you are good to go.
Any questions? Need a freelance coder for neural networks or android apps? Or interested in this kind of stuff? remko66 (-at-) gmail.com
No model is provided(oops...uploaded a bad one..please delete model.h5, it has only 5 minutes of training!)....it learns fast enough or your CPU...Just have a coffee or something and watch progress!
To start from scratchs:
- Leave images in high low (as they can not be generated yet from games description)
- Empty don't and flap (well leave one image in both else it crashes...take on from low for flap one from high for don't...bogus images also oke)
- params -m Train or run (train of course...)
directories:
trainingstuff:
- dont images where it flapped before and crashed...don't flap is the game looks like this (leave one image here before start..didn't bother to check if empty)
- flap images where the game didn't flap and crash..flap if you see this! (leave one image here before start..didn't bother to check if empty)
- high (picures of crashes on the ceiling)
- low (picures of crashes on the ground)
shouldbehigh/shouldbelow saves pics here after crashing so can check the flap crash investigator
Copyright: Yes, you have the right to copy...If you do drop me an email and make my day, though that is just a friendly request.
Remko Weingarten
remko66 (-at-) gmail.com
Disclaimer
This work is highly based on the following repos:
https://github.com/yanpanlau/Keras-FlappyBird
https://github.com/yenchenlin/DeepLearningFlappyBird
http://edersantana.github.io/articles/keras_rl/