#aiide

Imagine an AI-powered coding environment that helps you write code, debug errors, and build complete applications faster than ever. Tools like this are changing the future of software development.

Whether you're a beginner learning to code or an experienced developer looking to boost productivity, AI-assisted IDEs are becoming increasingly important.

AI in Video Games

C O N T E N T S:

KEY TOPICS

It turns out that as video game graphics have gotten better, the hardware used to produce them is increasingly well-suited to powering the AI future envisioned by companies like Google and Facebook.(More…)

Don’t forget that closely coupled with the challenge of general video game playing is the challenge of general video game generation, where plenty of other types of…

My excuse for the joint post? Sleep is important, even (or maybe especially) at conferences. I try to hit as many of the talks as I possibly can, but honestly? So much of the value comes out of the discussions that happen later that I do try to make sure I get rest. I missed most of the 3rd set of posters, as well as the Workshop reports to rest up as well. Either way, second verse, same as the first:

Keynote: Mark Walsh (“bringing Pixar empathy and character to AI”)

Today in “Talks that I wish I had a photographic memory so that I could talk about all of the details and nuance”: This talk. Unfortunately, due to this being a talk that HEAVILY leveraged copyrighted content, it’s not going to make its way to the internet (unlike the other keynotes). Which is disappointing, because this was a fantastic high-speed overview of concepts that go into making animated (and maybe possibly generated) characters not only believable, but enjoyable to watch. Not only did Mark Walsh go over the layers of character choices that went into designing the motions for the characters, but he also acted them out. Easily one of my favorite talks of the entire conference, and that’s even without this model of how to design the character motion (start with the animation basics, add a layer of intentionality [show the audience what the character is thinking] and then, finally, add the layer of characterization on the top of that).

Session: Strategy

Staying Hidden: An Analysis of Hiding Strategies in a 2D Level with Occlusions by Navjot Singh and Clark Verbrugge

Rock, Paper, StarCraft: Strategy Selection in RTS Games by Anderson Tavares, Hector Azpúrua, Amanda Santos and Luiz Chaimowicz

Implementation of an Automated Fire Support Planner by Byron R. Harder, Imre Balogh and Chris Darken

The first paper in this session was the only paper about stealth games that I saw that the conference this year, and it definitely took a different approach than I expected. They tried to implement a greedy agent for staying in shadows based on some particular metrics, to only mixed success. Never the less, I’m looking forward to seeing future advancements done, because Solid Robot Snake would be cool.

Next was a paper that took the top handful of bots in Starcraft, and then turned the game into a game theory problem: Clearly, some bots are laser focused at beating other bots, and if we have an idea of what kind of strategy our opponent is using, we can beat them by simply picking a counter strategy. So, the authors tried out a number of different strategies for picking a strategy from the pool, in essence creating a probabilistic Rock-Paper-Scissors game. As we’ll see in the main AI competition, this sort of result was used on the game level by the agent that won this year, so I’m hopeful that next year, we’ll see even more agents that try to change between strategies. The authors also entered the bot that created here, named MegaBot, into the Starcraft AI competition.

Finally, we had a paper by the literal United States Military, about planning strategies for providing covering fire for troops who were attacking a particular defensive position. The tactical concepts were a little lost on me, but apparently this was some-what of an open problem that needed work done. What actually struck me was that the approach taken actually produced some emergent behavior that the authors recognized as valid, tactically.

Poster Spotlights

There were a lot of posters.... I don’t have much to say about them in general, since they were delivered in a 5-minute shotgun style, and so I barely had time to jot down my impressions on one before the next person was up and presenting. So, in the interest of not lying about the posters that I didn’t get around to talking to at the poster session, I’m going to leave this particular section blank. (If anyone has commentary that they’d like to add, my ask box is always open!)

Keynote: Kevin Dill on Avoiding Artificial Stupidity

The main take away I got from this keynote was that even the most advanced AI systems that we have are going to never quite measure up to actual human performance, so a certain amount of smoke and mirrors is necessary in experience design to keep that under wraps. Which is really something to keep in mind when using AI, especially as an opponent, in experience design. To quote the speaker, the point of an AI opponent in a game isn’t to have the opponent win, but to lose with style (and, depending, to challenge the player). Both of these are important to keep in mind, but I do wish there had been more in the way of concrete examples of how he achieved these particular goals. Regardless, it provided some things to either think about or as a refresher of key experience design points.

Starcraft Competition Report

First of all, we have the results here. The top ranking bot, as I mentioned earlier, was a Terran player that build their play strategy around adaptive play, focusing on a building an army that was comprised of percentages of various units - mixing between, IIRC, Marines, Vultures and Siege Tanks. What makes this bot different from other bots is that it takes an agent based approach to decision making. Different experts within the system watch play, and pass up the priority that whatever subsystem that they control needs to be focused on. This allows the AI to multitask effectively, but also to switch unit production to counter the opponents build.

Of course, the one other bot that it didn’t do well against? The all-in 4pool build.

Demos and Playable Experiences

The demos and playable experiences were all super cool! We had:

a version of the Mixed-Reality Lemmings from the EXAG talk

Rogue Process, a Cyberpunk Rogue-like-like from Michael Cook

Elsinore, a game where you play Ophelia stuck in a Groundhog Day loop that resets at right about Act 2 of Hamlet

Some Conceptually Blended Mario levels, from a talk from the previous day of AIIDE

And Bad News, a game which deserves more explanation than I can afford to spend here.

I had a chance to play all of them, with the exception of Elsinore, which I instead just watched a number of people play, and I highly recommend keeping every single one on your radar.

Keynote: Simon Clavet - Motion Matching for Realistic Animation in “For Honor”

This was a really cool talk. I’m honestly just blown away with what we can do with computers these days. “Wait. You mean, we can just do a boatload of motion capture, and then just search over the poses when we want to switch animations? And we can do that in real time? Are you sure you aren’t trying to buy my soul?” This talk was just super fascinating on that level and the results are really quite stellar - the different animations (with a little bit of cheating at times because, let’s face it, we need it) blended together very cleanly, and I’m honestly impressed with the state of the art with Motion Capture if this is representative.

Session: Machine Learning

Demonstration-Based Training of Non-Player Character Tactical Behaviors by John Drake, Alla Safonova and Maxim Likhachev

Mystical Tutor: A Magic: The Gathering Design Assistant via Denoising Sequence-to-Sequence Learning by Adam Summerville and Michael Mateas

Portfolio Online Evolution in Starcraft by Che Wang, Pan Chen, Yuanda Li, Christoffer Holmgård and Julian Togelius

This was the final round of papers for the conference, and it ended on a high note. The first talk showed a way for an NPC to follow a path set either by a designer (or a player) in a way that allowed the NPC to both intelligently follow the lain-out path, but also to learn particular actions to take (sneaking, unlocking doors, etc.) as it moved. Honestly, I was super inspired by this paper as it provided a demonstration of what appeared to be a mechanic to build an entire game around (or at least a way to make escort missions less painful to play through). The fact that the authors used a Skyrim mod as a test case was just icing on the cake.

Next, we had a paper that literally threw every single M:tG card ever into a LSTM (a class of Neural Net) and then used it to generate new cards - and this works out surprisingly well! The secret sauce? Entering in a whole bunch of duplicate versions of the cards - but having each duplicate only be a distinct fraction of the initial card. The designed cards do feature some flaws (such as a creature that can go infinite all by itself) but this is some of the best generation seen to date.

Finally, we have a paper about building up a portfolio system for maneuvering individually units during combat in Starcraft - a skill that’s vitally important in the AI competition, since good micro wins combat. The approach works well, as the video demonstrations clearly show.

So, today was the first day of AIIDE-proper! Which means that the room changed slightly (ie the space doubled), more people were around (although some left) and more amazing talks!

Unfortunately, I didn’t have the mental stamina (or attention span, if I’m being perfectly honest) to stick around for the entire track today. That said? The parts I was around for were SUPER interesting, including both of today’s keynotes.

So, without further ado, the breakdowns:

Keynote #1: Katja Hofmann talking about Malmo - a Minecraft mod designed for plugging in AIs

This talk was super cool - it’s nice to see that Microsoft is really leveraging the power of Minecraft now that they own it, and it’s even cooler to see that they released this code into the wild. Personally, I found the actual use-cases shown to be a little... unrepresentative of what cool things you could actually pull off with the system (the agent never removed or placed blocks in the demos shown) but overall? I think it’s got a lot of potential and I’d love to see where it goes and what people do with it.

Predicting Proppian Narrative Functions from Stories in Natural Language by Josep Valls-Vargas, Jichen Zhu and Santiago Ontañón

PlotShot: Generating Discourse-constrained Stories around Photos by Rogelio E. Cardona-Rivera and Boyang Li

Fast and Diverse Narrative Planning through Novelty Pruning by Rachelyn Farrell and Stephen G. Ware

Game Level Generation from Gameplay Video by Matthew Guzdial and Mark Riedl

Co-creative Drawing Agent with Object Recognition by Nicholas Davis, Chih-Pin Hsiao, Kunwar Yashraj Singh & Brian Magerko

All of these talks were super interesting! First, we had a group that did machine learning on prose to try to predict the short of Proppian breakdown that the parts of the narrative featured - an interesting approach to be sure, but hedged in some talk about how Propp isn’t really considered the highlight of things. Still, the system does do pretty well, especially on a domain where the notion of a “right answer” is pretty vague. 

Next, the PlotShot system was super cool - taking Pictures and trying to build out stories that could connect the various images. Of course, this sort of system requires what amounts to a crapton of knowledge engineering and, of course, to do some image recognition to the images, but when that’s done? The planner actually does some really cool things.

I do find their notion of style (as an integer that depicts how aesthetically pleasing an image is) annoyingly reductive, but that’s just because I’ve got my own focus.

After was the Novelty Pruning system, which is a system where I need to spend time digging into the paper to really understand what they’re doing. Regardless, they manage to ignore less novel states to produce MASSIVE speedups in how they’re searching through trees to do story planning. I think. Like I said, really impressive work, but work that I need to read more into. 

Next was Yet Another Mario paper from Matthew Guzdial - talking about how he learned information from video footage, which was super cool. After pulling in a whole bunch of information from a lot of different videos, he managed to typify various parts of Mario levels, leading to some very cool designs!

Finally, we had a system on an AI drawing partner - or a “Co-creative Drawing Agent” if you’re titling an academic paper. This was a really interesting attempt at figuring out how to have an AI collaborate with humans on drawing projects, and it lead to a lot of questions about what that collaboration might look like - in addition to reinforcing a lot of what we already know about what humans want out of AI partners. 

Keynote #2: Jonathan Blow talking about what game design AI should incorporate next

In summary, this massive twitter thread. Seriously, go look at it - there’s a lot of great discussion happening in there. I think I’m going to have to go back to write a full post on this talk to really do it justice - there were a lot of interesting ideas that tie into a lot of the themes I was seeing throughout the conference as a whole.

Just like yesterday, we had seven different paper talks, each grouped broadly into 3 different categories. Unlike yesterday, two of the presentations were delivered via 

Procedural Content Generation

Mixed Reality and Procedural Content Generation in Video (Sasha Azad, Carl Saldanha, Cheng Hann Gan and Mark Riedl)

Intelligent Physiotherapy Through Procedural Content Generation (Tommy Thompson and Shabnam Esfahlani)

I’ll be honest, calling these papers simply PCG is doing them a bit of disservice - they, specifically, are PCG for Augmented/Virtual reality, and both are taking different approaches to the problem. While neither of these papers are necessarily in my domain persay, there’s a lot of really interesting game design questions that they both alternately raise and solve with their systems.

Theories

Proceduralist Readings, Procedurally (Chris Martens, Adam Summerville, Michael Mateas, Joseph Osborn, Sarah Harmon, Noah Wardrip-Fruin and Arnav Jhala)

Computatrum Personae: Toward a Role-Based Taxonomy of (Computationally Assisted) Performance(Benjamin Samuel, James Ryan, Adam Summerville, Michael Mateas and Noah Wardrip-Fruin)

The Proceduralist Readings paper was something I found super interesting - the notion of using ASP to automatically derive meaning from a game was never something that I would have thought up, but am super glad to see that it happened, and it gave me an excuse to think about how to ground out styles into meanings. The other paper was an interesting taxonomy of performative semi-computational systems, which was also super cool, even merely on the level of “wow I didn’t even know that many of these systems existed”

Platformers (aka Mario)

What Does Bach Have in Common withWorld 1-1: Automatic Platformer Gestalt Analysis (Johnathan Pagnutti)

Learning Player Tailored Content From Observation: Platformer Level Generation from Video Traces using LSTMs (Adam Summerville, Matthew Guzdial, Michael Mateas and Mark Riedl)

Deep Static and Dynamic Level Analysis: A Study on Infinite Mario (Matthew Guzdial, Nathan Sturtevant and Boyang Li)

And finally, this wouldn’t be a games conference if we didn’t have a whole boat load of papers about Mario. Again, not really my wheel house, but I definitely enjoyed all of the presentations.

Artificial Intelligence in Gaming

See up to 10 related YouTube videos at end. C O N T E N T S: KEY TOPICS Neuroscientist Jeffrey Lin wants to dramatically reduce people’s toxic behavior in online gaming communities, and he’s using artificial intelligence to do it.(More…) POSSIBLY USEFUL Responding to another question about when AI will reach human levels of intelligence, Hawking stressed that we don’t really know when that will…