Neural MMO – A Multiagent Game Environment
blog.openai.comMicrosoft project Malmo is attempting the same using Minecraft.
Github: https://github.com/Microsoft/malmo#getting-started
Homepage: https://www.microsoft.com/en-us/research/project/project-mal...
Hi! I'm well aware of Malmo -- it's a project I'm fond of, but not quite the same thing. I wouldn't quite consider Minecraft an MMO -- it doesn't really scale with the player base as nicely as, say, Runescape or WoW. There's a 2-pager on the high level objective and approach in the Github repo if you're interested.
What I find interesting about this is that the agents naturally become "pacifists". To optimize for survival, they try keep their neighbors at arms length and avoid conflict.
I wonder what would happen if they took the 8 species model that developed niches and introduced some kind of "boredom" feature. Or what might be more interesting is to have the agents produce offspring that stay with them early in life but consume equal resources and see if a "boredom" naturally occurs in the offspring.
I guess given that agents optimizing for survival creates routine, I'm curious where boredom or the desire to explore the world might originate from.
I wouldn’t see it as that surprising; most of human history is... not fighting. Its doing exactly that: staying at arms length, because war is expensive (and even victory can leave you weak enough to a third party invasion).
If you want to see combat, increasing the number of actors beyond sustainability will naturally cause fights until the same equilibrium is reached.. add total resource depletion (not just regeneration) to break the equilibrium again
You don’t need anything as complex as boredom to break pacifist behavior; just variable resource constraints will easily do the trick
I'd guess novelty seeking as a response to boredom is advantageous for learning and therefore being more adaptable to change.
It might be even more of a prerequisite though - babies play, bite, and shake stuff to learn about the world. Maybe without novelty seeking you can't get enough 'training data' to have a functioning generally intelligent neural net in the first place?
Yeah, the fitness metric was simply survival time, which is diminished as soon as you encounter conflict.
I would expect conflict behaviour to arise if the metric was modified so that say.. defeating another entity acquires the victor some portion of the other entity's current score (e.g. log(score(other-entity))).
Also, don't we forget all about (evolutionary) game theory?
There are loads of situations where pacifism, or a least tit-for-tat, represents a stable evolutionary equilibrium. This should not be surprising.
Plot twist: This research's purpose is covertly to make the best Runescape bots possible. :)
Why, there's not much money in those anymore unless you invest a ton of time on huge amounts of automation/infrastructure and at that point you couldve built a much more reliable SaaS business and playing RS isnt fun anymore.
source: invested a ton of time in that during HS when the getting was good ($5/m)
There is a lot of money in them, there have been plenty of companies and individuals making >$100,000 ARR by just working on MMO automation. Maybe some other SaaS ideas are more reliable, but it is more difficult to ensure you have a profitable idea and more uncertainty exists throughout the process.
Nope, but I did personally pick up the game again because of this project. It's an endless source of ideas.
maybe i'm wrong, but as far as movement is concerned it appears to be a 2d environment rendered to 3d.
edit: i'm misunderstanding the whole purpose of this, it's very cool.
You're right that it's 2d, but that's beside the point. You could say the same of Runescape -- the internal state is basically 2D and tile based. This is actually quite a nice feature, as it makes the environment very efficient to simulate.
That's really neat. I'm wondering if this can have implications for NPC's in MMO's proper.
I'm not looking forward to when they introduce chat though.
I look forward to the day where "scripted" characters can grow and learn. Characters with certain tendencies and traits set by writers can grow and learn in different ways since they are weighted differently and have different goals. Think the AI in S.T.A.L.K.E.R, but more "organic".
One thing I'm interested in is emergent language from scratch within a setting like this. Probably not the best idea to attempt to train a language model on MMO chat logs.
So far, they are replicating known results from evolutionary game theory (pacifism & niches) to economics (distance & diversification).
I wonder when and if they will surprise some novel results.
Well aware of a ton of results in older Alife work. A main goal here was to couple the ideas there with MMOs. I think a lot of older Alife projects puttered out because it was not clear where to take the environment next. In contrast, MMOs are an established game genre that we know how to build out and develop in order to support more and more complex play.