Hello there :) Video Playlist of both Surreal Numbers and Surreal Adversarial Learning, (And relation to Self-Supervised Learning): https://www.youtube.com/watch?v=gpIJ41IZj_o&list=PLKgB8_r-NpR5YcI1knBLHh7XZeh3ESKvn
Algorithm Summary: Become your own worst enemy, win, then win everywhere ;)
To Run it yourself: "05_DQN_Surreal.ipynb" (Should be good to go). This is just training a single agent on Cartpole with Surreal Adversarial Learning. Runtime is a few min, haven't optimized anything, (Not even sure if I was actually utilizing GPU now that I think of it... whoopsies). The entire experiment I ran consisted of running 6 runs for each algorithm, (18 total), and comparing their combined results. Ablation studies and MANY other experiments will happen. But. $. so...
Implications: BEEG YOSHI (https://www.youtube.com/watch?v=gpIJ41IZj_o&t=1s)
Paper coming, (I need money so... either have patience or give money haha)
I'll add the 2 other algorithm implementations later, (gotta make sure they ain't toooo ugly...) 1. normal DQN and 2. DQN with noise.