Show HN: Self-taught neural network learns to shoot a target | Hacker News Reader

Show HN: Self-taught neural network learns to shoot a target

github.com

60 pointsatum4721 comments

Am I the only one who thinks this is a little overengineered?

Not unlike most neural network solutions :-P

But hey, this seems to be a fun project more than anything else. And working in such low-dimensional spaces with an obvious disentangled representation (only shot strength and angle matter) might provide intuition to beginners.

yeah, I'm thinking about stealing wind strength and direction from the Worms game

Remember Gorillas from the QBasic language? That had wind speed and direction. Perhaps helpful.

https://github.com/GorillaStack/gorillas/blob/master/gorilla...

Gunbound was extremely complex with all kinds of variables including boomerang style projectiles.

probably not

think about it like this:

with a 10 minutes training the neural network takes about 8 shots to hit the target, so as the player, you have 8 shots to hit it before it hits you.

after 20 minutes of training the nn would take about 4 shots to hit you... so on and so forth.

the original idea was to have several levels of difficulty and distribute them across the game levels.

levels 1-10 poorly trained nn levels 2-20 better trained nn levels 30 and above you got one shot before the nn hits you.

anyway, it was just an idea

Never just an idea. Remove the "just". It is a realised idea and it was important that you execute the idea and, to whatever useful extent, finish it. Don't allow that negative wordplay. Don't make light of this and think there are no consequences in keeping that diminishing, passivity inducing "just".

I'll remember that. Thanks

On point! So much harder to execute and create something real and tangible than pontificate.

This is the light that shines through cynicism - the light that makes people brave enough to post new stuff.

This poster is correct - “just” takes away from yourself, it’s an idea.

well its kind of a hello world-ish example within this domain so I am fine with it

It doesn't look like totally self taught though

"If the shot was to the left of the target, adjust the weights to the right and vice and versa. On this step I'm not saving the training data, as I don't care for a miss shot."

self taught meaning that I didn't provided any training data

The environment is the training data and you are forcing the NN to hit the target. It isn't self taught. No.

If it's done like that isn't it just an overly complex, not very good PID controller (with one or more of those letters missing)?

I wonder if it will teach itself to aim at the edges of the target zone, given that's where the "adjust to the right and vice versa" method will end up landing shots.

I'm curious about the title of the module 'Dejavu.js'. Is it from another project?

Dejavu is like the fourth iteration of a fully connected neural network I've been working on. it needs a LOT of work to be considered useful, but I like to play with it from time to time

So you learnt a polynomial of order 2.

Great job, I'm quite scared of the killer robots now.

Reply on news.ycombinator.com