search

tamuseanmiller / alpaca-stock-trading-bot

📈 Q-Learning and Sentiment Analysis Stock Trade Bot Powered by Alpaca
GNU General Public License v3.0
42 stars 13 forks source link

Train Algorithm #1

Closed tamuseanmiller closed 4 years ago

tamuseanmiller commented 4 years ago

The algorithm for the number of years isn't correct and runs for long periods of time.

tamuseanmiller commented 4 years ago

Updated with last commit

GitHubEmploy commented 4 years ago

Hey, I am using the train algo, and for me, it's taking a very long time for some reason. I was running it constantly for the last 72 hours and it's only done one episode. I am running this on a Tesla K80 for computing, so it should work really fast. IDK what happened.

tamuseanmiller commented 4 years ago

Hey, I am using the train algo, and for me, it's taking a very long time for some reason. I was running it constantly for the last 72 hours and it's only done one episode. I am running this on a Tesla K80 for computing, so it should work really fast. IDK what happened.

There are a few reasons for this. Where the largest is because I messed up the training algorithm and am working on the update for it in integration right now.

GitHubEmploy commented 4 years ago

Okay, cool thank you for letting me know. Please tell me when the update is finished.

tamuseanmiller commented 4 years ago

Okay, cool thank you for letting me know. Please tell me when the update is finished.

Just working on some final testing for it, the train algorithm is complete if you want to download it from integration. You can even test it for me, if you would like.

GitHubEmploy commented 4 years ago

Ok, cool, I am going to test it out right now, and give you the results

GitHubEmploy commented 4 years ago

args = docopt(doc) #Line 195, train.py, Version Integration This line gives an error saying that it requires a byte like an object, so I replaced the arguments with variables.

GitHubEmploy commented 4 years ago

Ok, so testing it the results are way better. Instead of 72 hours, it took me 10 minutes to train 1 episode! This works way better and I think that this is good so far. Also, the docopt error is also happening in the new version of your eval file so you need to fix that there too. Thank You.

GitHubEmploy commented 4 years ago

Also wondering, why do you force it to use CPU? Wouldnt GPU be faster?

tamuseanmiller commented 4 years ago

Also wondering, why do you force it to use CPU? Wouldnt GPU be faster?

Training in this way is actually faster using CPU according to the original author

tamuseanmiller commented 4 years ago

Ok, so testing it the results are way better. Instead of 72 hours, it took me 10 minutes to train 1 episode! This works way better and I think that this is good so far. Also, the docopt error is also happening in the new version of your eval file so you need to fix that there too. Thank You.

I was wondering, can you reproduce this docopt error? I can't seem to trigger it.