Kenneth Reid, The Factory Must Grow: Automation in Factorio

Kenneth Reid, MSU
11/20/20 BEACON Seminar
The Factory Must Grow: Automation in Factorio
Efficient optimisation of resources is paramount to success in many problems faced today. In the field of operational research the efficient scheduling of employees; packing of vans; routing of vehicles; logistics of airlines and transport of materials can be the difference between profit and infeasibility. In this presentation I will provide an overview of the video game Factorio and present the analogous problems and potential solutions. Additionally, I will present the interface we have developed to allow external optimisers to communicate with Factorio, and explain some of the fitness functions which define a good solution from an ill suited one.

[Trupen]

This is very interesting!
Thanks for working on this project :)

[siriguillo]

Most accurate Factorio review yet, 10/10 would watch again

[KenReid]

Thanks for posting this, and for the opportunity to present our preliminary work.

If anyone has any further questions, please reply to this comment, and I'll respond as soon as I can.

[roadtrain_]

Ah yes. The factory must grow to sustain the growth of the factory.

[Brandon-rcvp]

This conversation around 18:00 about human-AI interaction is extremely interesting when considering its application to Factorio, I suspect there will soon be a mod that does exactly, thus bringing GPU utilization in line with CPU utilization, further improving the efficiency of the Factory-Machine-Human interface for the ultimate optimization.

[mxdanger]

I really enjoyed this video presentation.

[ether]

this was excellent, the person who said you should start a youtube channel about your progress is very right! I'm no academic but i've been thinking about the education part of it. Factorio has taught me and many other people how to properly signal trains, something that usually is a very expensive and stressful course yet i've learned it while enjoying myself and probably in a shorter amount of time than those courses would take due to how interacting with games is both enjoyable and an actual hands on learning as opposed to a classroom setting. it's unfortunate that most educational games are generally suboptimal at both teaching and entertaining.

[ganis]

Now that's just spiffing. You are translating the thoughts of so many players into a more formalised framework. Good Job!

[theoaketree]

I'd very much like to see a highly optimized A.I. speed run Factorio. I can't even imagine how insane it would be to witness once they got to construction bots. Things would really ramp up at that point.

[palmberry]

The factory must grow
The factory must grow
The factory must grow
The factory must grow
The factory must grow

[mrJonnyH]

Might I suggest you contact people that already do in game optimization eg grey goo Projects
Or perfect splitters makers. Also to make your testing easier get a creative mod.

[Brandon-rcvp]

On the final question regarding sustainable economics, there are mods that introduce recycling mechanisms and waste products that could be expanded upon to create more accurate simulations allowing that type of question to be explored, it would also introduce a style of game play I have been looking for so yes please do this :)

[Turalcar]

A lot of great strategies come out of code analysis. One that comes to mind is the Super Mario World Credits Warp. (This is an arbitrary code execution bug that doesn't seem to trigger naturally and bugs are few and far between in Factorio). Another way this is often used is pseudo-RNG manipulation (Minecraft max diamond mining, Doom II Map 30 pacifist TAS). I don't know how far off AI is from being able to find these even if we include code or the entire memory state in it's inputs.

[DavidLindes]

20:22 - Well, I don't know how plausible it is for you to get these AIs to the point of launching a rocket. But if you do, I look forward to the new speedrun category to show us how the AI does it -- because I imagine the AIs will do it _much_ differently than how the humans do, because the constraints will be so very different.

[joshinils]

I think they could probably make it possible for a challenge to be played in the demo version of the game, or create a special version just for that

[Krazycat]

Hey, another approach to the solution is to start at the end. Set a goal of what you want accomplished. 1 science per second, one rocket per minute. Factorio is a game that as you play you need goals to accomplish, the base game is to launch a rocket to get off the planet. Which could be accomplished by skipping a lot of the other production of the game. But if you start with that and say add 1 science per minute. Then you add a lot more options to the game. I am not a coder, but I have many hours in the game and have a hard time getting to just launching the rocket still. I am sure I am over thinking the process to get there. But I thought that my thoughts and ideas would be a help to you and your group.

[jmvjeroen]

So, is there anyway we can keep track of progress?

[trondhelseth]

if you want the REAL factorio experience, pyanodons

[bigeteum]

I think if automation of factorio is done then we are close to a general ai... But I must say this is a very difficult problem to tackle, maybe require something like a genoma project type of effort.

[kingcrimson]

This is fascinating. I imagine the problem with a GA approach would be the sheer number of optimisation controls. Maybe using GA to solve subproblems would make more sense than a 'global' solver. What is the current status of this project.