Reinforcement Learning from scratch

How does Reinforcement Learning work? A short cartoon that intuitively explains this amazing machine learning approach, and how it was used in AlphaGo and ChatGPT.
Part 1 of 3.
0:00 - intro
0:13 - pong
0:28 - the policy
0:51 - policy as neural network
1:32 - supervised learning
2:51 - reinforcement learning using policy gradient
4:24 - minimizing error using gradient descent
4:45 - probabilistic policy
5:01 - pong from pixels
6:58 - visualizing learned weights
8:18 - pointer to Karpathy "pong from pixels" blogpost

[@darthvader4899]

this is video is super underrated. In fact the whole channel is underrated.

[@themathguy3149]

Your Channel IS SO GREAT, I share with all my eng friends for you to get more visibility!

[@metaljacket8102]

This is really awsome! It's the best video that explains DRL in such an easy to understand way!

[@tushargupta1999]

This video is amazing. You explained everything in such a simple manner. I am feeling really motivated to learn more about reinforcement learning and neural networks after watching this.

[@a.aspden]

Your videos are great. Looking forward to more!

[@ashketchum1244]

I don't know how I stumbled upon this video but that was very interesting and intuitive to understand. Thank you.

[@marcinstrzesak346]

Great video, very helpful, easy to understand.

[@gmjammin4367]

Amazing video as always :)!

[@moldo800]

Excellent. Congratulations ❤

[@mado.madeleine]

Super helpful! Thank you 🙏🏽

[@jameslibby5215]

Very very underrated channel

[@benc7910]

Underrated, two Rs

[@jameslibby5215]

@@benc7910 thank ya sir

[@nikbivation]

thank you for this!

[@cloudysh]

This was so surprisingly great :3

[@themax2go]

agi: 1. ai develops understanding of win-loss conditions and sets policy params (inputs & actions) accordingly. 2. ai creates (= designs & builds) training env(s). 3. ai iterates, avals & adjusts policy parameters accordingly 4. done (or validation run(s) w/ human(s))

[@ireoluwaTH]

Thank you!!!

[@mohajeramir]

Excellent

[@CptDoge-rn3ou]

I really like the way you visualize what you are talking about. Thank you for putting in the effort!

[@kniv0gaffel]

Brilliant

[@solveigberling1662]

That was dope

[@BlueBirdgg]

Can you playlist each one of your topics plz? I wanted to post on Twitter(X) your video topics but could only post a single video at a time. Great content by the way. Ty very much. Your perspective on some topics helped me a lot to get a more intuitive understanding.

[@g5min]

Good idea! Here's one on generative AI: kzbin.info/aero/PLWfDJ5nla8UoR8P7AGqVw7ZPjXajUFLMo Here's one on reinforcement learning kzbin.info/aero/PLWfDJ5nla8UoexEaLqVMw7q3Ft0vRYscL Here's one on LLMs + text-to-image kzbin.info/aero/PLWfDJ5nla8UoG2mvvHs_OS0asAKC5HJeu

[@BlueBirdgg]

@@g5min Ty!

[@edvinbeqari7551]

What is your reward function for the pong game? I did a similar pong game and I couldn't get it to learn.

[@bombur9007]

how many layers should such network have

[@nischalyou]

whats the name of this video game ?

[@mineq4967]

but by what number do you change the weights like you never told us

[@axe863]

Simple Reinforcement learning is extremely dangerous in certain nonstationary environments 😅

[@macratak]

ah yes, reinforcement learning. a fundamental computer graphics technology

[@g5min]

I think that character/game-AI is pretty central to graphics

[@pw7225]

Why so negative?

[@revimfadli4666]

​@@g5minespecially AI image generation or processing nowadays

[@FRANKONATOR123]

Can you share the source code for this project

[@g5min]

You can follow the link to the Karpathy site at the end of the video, repeated here: karpathy.github.io/2016/05/31/rl/

[@herikaniugu]

Imagine using reinforcement learning in quantitative finance 😊