AI Learns to Outrun Police Officers

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Video showcases AI trained using Deep Reinforcement Learning.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

artificial intelligence, ai, machine learning, ai learns, deep reinforcement learning, ai training, evolution, neural networks, cozmouz #ai
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

[SinnaMon-sp]

so basically the equivalent of putting a baby in a timeloop and teaching it to steal.... i approve

[anador]

I think there should have been a negative reward for jumping off too. It was clearly a preferable strategy to risking touching police officers, especially before discovering that coins give rewards or when the AI thought there were no way to get coins.

[fureyXD]

You should have added a negative reward for getting seen by the police, that way Loki will sneak around them instrad of speedrunning trough them, maybe a level where police couldn't be outrunned could have helped

[forcelightningcable]

I like how Loki figured that it’s better to die than get caught by the pigs

[grandpretredesalpagas]

if the police start using robot dogs, we will start making robots cat robber

[kitkat-bh]

i guess you could say hes _lowkey_ a fast learner

[Benw]

The problem with videos like this is that the AI can overfit to a specific map. You need to have some sort of shuffled dataset or randomly generated sequence of maps/coin arrangements for proper training.

[tach]

"That's her officers! That's the woman who programmed me for evil!" - Bender

[Pasu]

11:17 I think this happens because the AI only learned to effectively collect coins in the one direction, or gets confused by there being no police to dodge.
AI is not that good at changing its perspective, since it has no real correlation between x, y and z. It doesn't know that they are just sides of the same coin, it only knows what outputs will change them individually.
I saw a video of a table tennis AI that worked great for one player, but once they spun it around for the second player, it just fell over, because it only learned to stay upright while looking in one direction. Their solution was to rotate the coordinate system with it (rotating a parent object and using local coordinates probably).
I think something similar may work here too, by changing Loki's sensors to be relative to his orientation, thereby eliminating the need to correlate different axes (unless you are already doing that).

[Golden_Projects]

if you increased the reward from coins by dividing it by the amount of time from the last coin (less time more reward) you'd also make it so that he doesn't skip nearby coins to often, but it would also result in more speedrun-ish behavior

[redstonewolfx]

You might want to add a very small negative reward that accumulated over time, and/or a time limit, so Loki is encouraged to pick up the pace. He might also be less scared of the police, as the penalty for meandering aimlessly will eventually be worse than just running for it.

[kitsunemusicisfire]

Loki isn't evil he's just a silly guy

[emad.]

Programmers already teach AI how to do crimes. Perfect for our Sci fi apocalyptic fantasy doom.

[Digby]

Maybe we shouldn't be teaching AI to break the law, maybe that's just me.

[corruptedmineral]

damn i can finally create army of ai thief with ability to escape on its own

[L-ivlx]

"started to associate negatives with something tagged as police"

it started using twitter

[InksAutism]

He kept getting caught when teasing the cops

[a.j.outlaster]

This is cool, But wouldn't the A.I. learn more effectively if the levels scale slower in difficulty and repeated the same sort of scenarios?
Idk, This just seemed to scale at a rate that's fine for players but maybe staggering for an A.I.

[FlaiseSaffron]

In this video: programmer explains criminal psychology without realizing it.

[aurnok]

When Loki moves randomly he kinda looks like a speedrunner lol