Solving Atari RAM environments using causal entropic forces


#1

We have found a way to explore the state space of an AI problem that allows us to get high reward, and low probability trajectories without using any neural network. These are some examples that I calculated using my laptop.

Qbert

MsPacman

This is a totally new approach to AI based on the causal entropic forces principle. If you are interested in the code and theoretical framework supporting our work please contact us.


Atari game "Ms Pack Man" played without previous training
#2

New users can only use two links per post, so here you have some more examples. These two have less than 100 games played because it takes too much time on my laptop. Once we get an sponsor we will run it in AWS.

VideoPinball

Tennis


#3

Hello Guillem,

This looks pretty cool. Is the code available on Github?

Thanks,

Aidan


#4

Sounds interesting! Can you make your system available (i.e. GitHub) or send it to me?


#5

Looks great. Documentation or source available anywhere?


#6

We will make it public in some months, we want to beat all of them first :wink: