Subsumption Architecture, AI and Academics
Sorry for the misleading title – this won’t be a deep look into AI models, and neither will it link to any brilliant new papers concerning AI architecture. There will also be no diagrams and no pseudo source code. All you’re getting is a musing on differences between the academic and the commercial world. I’m sure you’ve read countless of articles about that issue already, but I have something new: I have anecdotal evidence to spice things up! Cue the crowd’s cheer.
The time: February 3, noonish.
The place: A stinky bus full of tired and half-drunk travellers going from France to Sweden; at this time the bus happened to be located in Denmark, but that’s highly irrelevant.
The cast: A mix of people who have studied Computer Science but ended up on different routes in life: a PhD student/consultant, a PhD student/biologist, a programmer, and a beer-swilling bastard. I only added the last bit for the amusing alliteration, but I’m sure you can guess whom I’m referring to. The main point is that two chose a rather more academic route than the other two.
I had brought along the latest (or by now old) issue of Game Developer on the journey, as some light reading compared to Michael Moorcock’s Wizardry and Wild Romance: A Study of Epic Fantasy. No, the light reading part wasn’t ironic at all – Moorcock’s book is excellent and he is totally dripping with wit, but the constantly changing writing styles (since he quotes different authors) makes it a non-trivial read for a poor Swede. Either way, in the Game Developer mag there was an article about AI models for games, and one of the more academic persons happened to browse through said article. I’ll make an attempt to describe an exchange of words that occurred around this time:
“So, was there anything worthwhile in the magazine?”
“Nah, not really. Interesting to read about the AI model, though. They’re acting like this is the next big thing in game development research, but it’s really just a simplified version of subsumption architecture. It’s been common knowledge in AI research since the 80s!”
I was pretty sceptical at this reaction. Sure, game developers don’t really keep up with academic research, but there had to be something new! Unfortunately I hadn’t – and still haven’t – read the article, so I can’t say anything for sure. The argument briefly touched the notion of game developers being result-centered and academics focusing on demos and theory rather than practical applications. I advocated this viewpoint, but I was met with arguments that the dude’s PhD work had included active attempts to develop AI models that could be practically used in games, and that he had worked together with industry representatives on this matter.
Then we arrived at the ferry from Denmark to Sweden, and had to get off the bus; so the discussion stopped.
I remain slightly sceptical, though. I have a nagging gut feeling that the viewpoints of a game developer and a person with a more thorough academic background simply are too separate to merge that easily. I’m convinced that said PhD student acheived excellent results that would be practically useful in game development…but I’m also convinced that despite all his efforts, all that it has resulted in is a demo; a proof of concept. An academic proof of concept that can be used for homebrew development, but probably won’t get incorporated into commercial game development projects.
I also have a nagging feeling that the PhD student in question focuses on potential. He sees what his projects can evolve into – he sees the possibilities for evolution that a good, solid theoretical foundation has. But I also suspect that a commercially-inclined game developer would only look at the actual output. If the AI he beheld didn’t produce effects that were wildly superior to existing simpler models, he probably wouldn’t care much for the potential in the project. This might be due to a lack of understanding of the potential, or simply a sound business sense since old-fashioned methods are cheaper and yield good profits either way.
Of course, this is a vague generalization. There are exceptions in both academia and commercial game production, and I have no data to support my claims in any way. All I have is this gut feeling. It would be very interesting to dig up a bunch of academic AI projects and interview professional game developers about their spontaneous opinions on what these projects could be used for.
All names have been omitted from this article to protect the innocent. Me, that is, because Peter would do harmful things to me for paraphrasing his words incorrectly.