Author | Message |
---|---|
rabbott
Posts: 1649
|
Posted 22:33 Oct 12, 2018 |
The previous post discussed how to install the OpenAI Atari system on your computer. It worked on mine, but I couldn't get it to work on the computer in ET-A129. (I don't know why.) Try it, and let me know if you can install and run it. Since then I've decoded the environment. Here's a little program that will let you play interactively. The system will prompt you for an action. (0, and 1 do nothing; 2 and 4 move up; 3 and 5 move down.) A nice thing about playing this way is that there is no time pressure. The game pauses until you enter a move. You will find that there is a fair amount of noise. The commands don't respond immediately, and they don't always move the same distance. So it's a harder game to play than it looks--even when there is no time pressure. The program also decodes the observation. After every user action, it prints something like the following:
This is a list of the blocks on the screen. Normally there will be three blocks: the two paddles and the puck. (At the start for the first few actions, the computer's paddle and the puck are not yet on the screen.) Each block is shown as ((row_top, row_bottom), (col_left, col_right)). In the example above, the first block (the computer's paddle) occupies rows 117 - 132 and cols 6 - 9. The second (smaller) block is the puck. The last block is the player's paddle. The blocks are always listed in this order. (The numbers are somewhat relative since I cut off some of the screen which isn't of any use when deciding on a move. All you really care about are the relative positions.) If anyone can get it running before tomorrow, we can look at your implementation. If you are really ambitious, you might write a program to play automatically. (This will not be reinforcement learning. It will be more like solving the cart-pole problem once you know what the objective is and what the observation numbers mean.)
Last edited by rabbott at
10:46 Oct 13, 2018.
|
jpatel77
Posts: 44
|
Posted 01:28 Oct 13, 2018 |
I faced an issue while installing the atari_py using the pip command. The error I got was : Could not find a version that satisfies the requirement atari_py (from versions: ) So, I cloned the repo and I tried to install it from the local dir. git clone https://github.com/Kojoley/atari-py But still received some error related to gcc while installing atari_py, So I pip installed gym: pip install gym After that, script seems to be working fine. Not sure if this is what is intended but the interactive agent script is also working fine. Still not able to install atari_py from that git though. |
jungsoolim
Posts: 38
|
Posted 19:26 Oct 14, 2018 |
I had to do few things to install atari-py on Windows 10. 1. Logged in as an administrator. - (Depending on OS settings, your account may need administrator rights to install a python library.) 2. Installed C++ compiler. MS Visual Studio IDE will provide C++ compiler. I installed 2015. It is free and available at https://visualstudio.microsoft.com/vs/older-downloads/ 3. Changed the project python interpreter to point the python interpreter which has atari-py. For example, I installed Anaconda, python37 and python36 in my computer. In default, atari-py was installed on python36 library. The project python interpreter can be set by selecting File => Settings => Project: studentWorld Project => Project Interpreter => choose the python.exe which has atari-py Finally, interactive_Atari.py runs and I can play the game! |