Whitney AI
Would you like to react to this message? Create an account in a few clicks or log in to continue.
Whitney AI

Homebase for Whitney's AI Class


Jason Jewik

2 posters
Whitney AI :: Current Projects :: Everyone's Projects
Go down
avatar
Jason Jewik
Name : Jason Jewik
Posts : 14
Join date : 2019-01-15
http://jasonjewik.com
+

Jason Jewik Empty Jason Jewik

Tue Jan 15, 2019 8:37 am
Current Projects

  1. Whitney Code Jam: February 16 Schedule and Assignments
  2. Whitney Code Jam reflection Post-Mortem Report
  3. Training a neural network to play Penguin Run: April 5

    • Create an autoencoder convolutional neural network (CNN) that can pick out important features of the screen (March 20)
    • Create a recurrent neural network (maybe a Long Short-Term Memory Network (LSTM)) that takes the output of the CNN and can output button presses
    • Train the LSTM by letting it play the game
    • Modify the LSTM's inputs (what portion of the screen to look at, how many layers, etc.) until the AI can play the game at least as good as me (high score of 1000 pts)




Last edited by Jason Jewik on Thu Mar 14, 2019 8:26 am; edited 6 times in total
avatar
LeeHammmm
Admin
Name : Liam Abalos
Posts : 20
Join date : 2019-01-14
Age : 23
+

Jason Jewik Empty Re: Jason Jewik

Tue Jan 15, 2019 8:41 am
Is this Code Jam supposed to be the Hack-A-Thon you've been working on?
avatar
Jason Jewik
Name : Jason Jewik
Posts : 14
Join date : 2019-01-15
http://jasonjewik.com
+

Jason Jewik Empty Re: Jason Jewik

Tue Jan 15, 2019 8:53 am
Yeah, the name we settled on was "Whitney Code Jam" because we didn't like the possibility of the term "hackathon" sounding too daunting for beginners.
avatar
Jason Jewik
Name : Jason Jewik
Posts : 14
Join date : 2019-01-15
http://jasonjewik.com
+

Jason Jewik Empty Re: Jason Jewik

Tue Apr 02, 2019 3:56 pm
Progress so far is that I've figured out how to get training data for the network. The program takes a screenshot of the game window, scales down that screenshot to 5% the original resolution, then applies a black-and-white filter. On average, this process takes 0.05 seconds per screenshot, meaning the program isn't really capturing every frame, but it's still getting a good deal of them. While that's running, the program also listens for key presses. On any given frame, if I press the space bar, the program will write out "1" to the inputs array; otherwise, it'll write out "0". To save data for future training, the program can store them as .txt files.

Although I plan to feed all this training data to the network as Numpy arrays, I thought it'd be cool to see what the network is "seeing", so I put together some of the reconstructed images and made them into videos.

Video 1
Video 2
avatar
Jason Jewik
Name : Jason Jewik
Posts : 14
Join date : 2019-01-15
http://jasonjewik.com
+

Jason Jewik Empty Re: Jason Jewik

Sun May 12, 2019 1:58 pm
I finished creating the PenguinFlow thing and made a Google Slides presentation about it.
Sponsored content
+

Jason Jewik Empty Re: Jason Jewik

Back to top
Permissions in this forum:
You cannot reply to topics in this forum