Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[§k]

Cockroaches appear to use just two pieces of information to decide where to go, namely how dark it is and how many other cockroaches there are. A study used specially-scented roach-sized robots that appear to the roaches as real to demonstrate that once there are enough insects in a place to form a critical mass, the roaches accepted the collective decision on where to hide, even if this was an unusually light place.[39]

How they show intelligence with seemingly simple rules is very interesting.

I followed the python roguelike guide when I started to make my first game last month. I wonder if there is any easy to understand machine learning guide, or is it really rocket science that needs at least a college or online course?

[Cathar]

What would be the cost of letting a monster machine-learn some basic roguelike tricks from scratch, say, kiting, pillar dancing, retreating to corridor when being surrounded?

Well, I don't know, I just follow the professional starcraft scene (from afar) and Alphastar was a big event, especially because they anihilated the human players they were pitted against. 5-0 against two pro players.

I assume what you want to do would be harder than programming alphastar tho. Alphastar plays against itself ; protoss vs protoss in a symetrical map. Both players have access to the same ressources, which means you can train agents against each others.

If you want to train an agent to play against something that functions according to different rules (eg an rts agent vs a rpg player) you can only train the agent against human players, and not against itself. Human input becomes a necessary step in its training process, you can't just make a thousand tournaments every hours like they did for Alphastar.

[Dorsidwarf]

That kobold game is masterfully infuriating and not in the slightest bit fun. Really puts you in the headspace of the character too

[§k]

I assume what you want to do would be harder than programming alphastar tho. Alphastar plays against itself ; protoss vs protoss in a symetrical map. Both players have access to the same ressources, which means you can train agents against each others.

I want to let the player be at the same power level as the majority of monsters. And I don't actually want the monsters to beat human, just want them to be moderately smart and somewhat unpredictable.

I hope that the scale would be reasonable for hobbyist developing.

[§k]

That kobold game is masterfully infuriating and not in the slightest bit fun. Really puts you in the headspace of the character too

At least it's fun for a while...

After finding out that I couldn't possibly defeat that legion of kobold in their tavern, I made a suicidal charge and ate their apples and sausages and crushed their squeaking baby kobolds, hahah...

[Egan_BW]

If you want to train an agent to play against something that functions according to different rules (eg an rts agent vs a rpg player) you can only train the agent against human players, and not against itself. Human input becomes a necessary step in its training process, you can't just make a thousand tournaments every hours like they did for Alphastar.

I see no reason why you couldn't make a separate "player-character" AI in addition to your "monster" AI and have them train on each other. It would take more work to set up, but wouldn't train any slower than an AI that trains against another version of itself.

As a side effect, you now have AIs able to take on either role, meaning that the human player can take either role as well.

The main problem I see with applying this kind of AI to games is really the fact that you probably want to have several of them running at once on a single computer, which is going to be just someone's PC, not some kind of supercomputer. It's going to be hard on the CPU.

[Starver]

with select 'roaches given player-like abilities to use and exhibit,

(Exactly how well you could do that might vary. Train your' PC AIs to be the best they can against the NPCs, simultaneously, of course, but hope you don't end up with an Pink vs Orange AI arms race where you're trying to get a Red vs Blue one, with totally unNPC and unPC-like behaviours. Though that might just produce your breakthrough game-concept that takes the world by storm, like infinitely varied Candy Crush levels in a world previously looking for Grand Theft Auto scensrios.)