ai plays super mario bros for 5 minutes straight

i am in a writers block at the moment so no music video for now
i present another one of my hobbies (AI if it isn't obvious)
this is a reinforcement learning network using the A2C (Advantage Actor-Critic) algorithm to play 1-1 in super mario bros
in the video the episodes presented are around episodes 60-90 in training
the ai receives positive reward for any actions that involve getting to the end of the level (moving right and squashing enemies) and receives a consequence for actions that don't involve getting to the goal (being alive, moving left, and dying)
layman's terms:
be good at the game = 👏 🎉
be bad at the game = ❌ 👎

[@Datceu2]

Bro forgot to hold the button

[@out-of-will]

yes even at episode 1080 it still does not jump the pipes first try (but gets unstuck quicker or jumps between pipes entirely)

[@out-of-will]

parts where it freezes are where both agent and critic are evaluated for loss

[@aRIE_vcr]

😭😭 do you have a platform where we can be friends? I need more composer buddies

[@out-of-will]

youtube will have to do for now until i set up a discord server i'll check once every 5 years