DeepMind Lab2D(github.com)

49 pointsby martinlaz6 лет назад8 comments

https://github.com/deepmind/lab2d

8 comments

programd6 лет назад

I've taken a look at the API for Lab2D and it's very abstract. What you have to work with are functions for strings, integers and floating point numbers glued together with callback functions and a bare-bones engine for running all of it through the equivelent of a game loop. In effect you have to define the meaning behind all those numbers and callback functions to create your own simulation. There are no pre-defined humans, robots, vehicles, generators, conveyer belts, minerals, etc for you to interact with.

It's all very hardcore and suitable for numerical academic research, but IMHO not very accessible to a what I think is a fairly large audience which wants a ready made world in which to play with some AI agents. At least I assume that such an audience exists based on the many posts here mentioning personal AI projects.

What I think is needed the equivelent of a game engine with an API allowing you to plug in AI models in place of the entities within it. Something similar to world games like Factorio, Rimworld, Hearts of Iron, Anno 1800, but with an API allowing you to create and control in-world assets. Something similar to the StarCraft API used by AlphaStar.

I think we really need a standardized world sim as a server, equipped with some gRPC API to control its entities and get out information. You run your world sim on some machine or Kube cluster and interact with it via a well known API. Plug in whatever AI you need into the API spawn a couple million agents on a couple million tile world and off you go. I suspect you could do a very fast sim with modern multi-core desktop CPUs.

Hmm, I am looking for an interesting project...

krasi06 лет назад

You may be interested in BWAPI. An API for developing StarCraft 1 BroodWar AI. The community around it is amazing and there are a few yearly competitions. The current top bots are very strong already.

https://github.com/bwapi/bwapi

mlyle6 лет назад

There's always crobots (c. 1985) / netrobots

https://github.com/netgroup/netrobots

29athrowaway6 лет назад

It would be cool to have a UI like this.

https://otoro.net/slimevolley/

befeltingu6 лет назад

That looks pretty slick...

I dont know how much having a nice UI helps with understanding / training but it sure makes it nicer to work with and helps in demonstrating concepts.

jan_Inkepa6 лет назад

Nice! Looks like it even does level generation. I’ll have to give this a closer look.

I made a medium-popular scripting language/game-engine ( https://www.puzzlescript.net/ ) whose games are by chance in the domain of this framework. I wonder if it’d be worth it to try write a wrapper for it...

ipnon6 лет назад

It seems like creating an API for your agent to train with can be as difficult an engineering problem as the machine learning itself.

awal26 лет назад

The fundamental truth of ML at the moment seems to be that gathering data takes at least as much effort and infrastructure (and often much more) than actually training the models.