Tobold's Blog
Thursday, March 30, 2023
AI and the curse of the MMO

Once upon a time this used to be a MMO blog. The era of the classic MMORPG came and went, and while several games are still technically alive, it isn't a hot genre anymore. However, many of the games that came since then are still technically "MMO" games, that is to say that they are online games with large numbers of players playing together. The whole games as a service or live service games idea isn't fundamentally different from MMOs; it is mostly the gameplay that varies, not the underlying structure. Unfortunately that means that they inherited the same curse: Network effects going both ways.

Imagine you would be the only person on a Fortnite server. It would be a pretty boring game. Players in these games are not only the customers, they are also the content for other players. The more people play, the more interesting the game becomes. That leads to a network effect, where a game becomes popular shortly after release and has rapidly rising user numbers. The game attracts players because it is growing. But that works both ways: After some months or years, the game reaches its peak, then it starts declining. And the decline is as self-accelerating as the growth was. There is typically a die hard base of around 5% that sticks around until the game actually closes, but the other 95% are tourists that can disappear as quickly as they had arrived. History repeats itself, the curse of the MMO strikes again, even if the games don't call themselves MMOs anymore.

Live service games aren't doing so well lately. Not only were there too many of them, which accelerates the tourist problem described above. But games also reached some sort of a plateau regarding graphics. Over the last decade a lot of these games sold on the premise of being graphically more stunning than their predecessors. Now a level has been reached by the average game that is hard to beat at a reasonable cost. Other genres of games, in which graphics aren't quite as important, have reached that plateau years ago. You can't visually distinguish for example Octopath Traveler from Octopath Traveler II, despite 5 years of difference in age. Age of Wonders 4 coming out in May this year is a lot prettier than Age of Wonders 3 from 2014, but most of the difference had already been reached by Age of Wonders: Planetfall in 2019.

One of the hot topics of 2023 are the advances in Artificial Intelligence. Machines are getting better at pretending to be humans, creating text and images of a higher quality than the average person. Sure, if you look twice you'll notice the AI is having problems getting the number of fingers on a human hand right. AI chat bots sometimes get basic facts wrong. But Artificial General Intelligence is a much harder problem than the application of artificial intelligence to specific tasks or problems. Computers have gotten quite good at beating humans in games like chess or go, well beyond the capabilities of average players.

And there maybe is a chance for gaming: Instead of pouring more money into graphics that are getting exponentially more expensive for little gains, why not improve the AI of games, which has been a widely neglected field in game design? There have been a lot of lame excuses for that, like saying that if you made a good AI, it would always beat human players, making the game less attractive. That is just a fundamental misunderstanding of the role that an AI would play in a game: The goal is not to make an AI that can beat players reliably, but, like with chat and image creation applications, to create an AI that can pretend to be human at first glance. You can buy chess computers for under $100, and they can easily be adjusted to whatever level their human opponent is playing, as long as that human isn't a master chess player.

Anyone who has ever played the same game both against the computer and against human players knows that the two feel very different. Current algorithms that play computer game are often highly predictable, because they are simplistic and don't use any actual AI technologies. If you search for "advanced strategies" for single-player games, these almost always consist of knowing what the computer opponent will do in a specific situation, and exploiting that foreknowledge. The interest of playing against real humans comes from them being less predictable. It wouldn't be so hard to create game AIs that are less predictable, or even have a bit of machine learning features, changing their strategy if the human always plays the same way.

Let's face it, the people you meet in a massively multiplayer online game or live service game or multiplayer strategy game aren't geniuses either. The average person you meet, by definition, has average capabilities. You could create bots that have a similar level of game performance, similar level of unpredictability, and a much lower level of obnoxious and trolling behavior. Which means that you could create a live service game with a constant number of participants, but varying percentage of bots. You wouldn't get that "oh, this games feels empty, I'm leaving" negative network effect to the same level anymore. AI could break the curse of the MMO, just by being good at pretending to be random average people, not by being exceptionally good at playing the game.

Strategy games would probably benefit even more from better, more human-like AI than live service shooter games. Games are longer, which leads to frequent problems with real human opponents having to leave for real-life reasons or rage-quitting when it becomes obvious that they are in a losing position. Many strategy games offer both single-player and multi-player options, but the games in which multi-player is the more popular option are usually a lot faster and simpler games, like Among Us. Getting a more human-like AI opponent for longer games would be quite a boon.

There has been a lot of speculation, frequently exaggerated, about what the progress in AI means. Some people invoked the rise of the machines, or at least making a lot of current jobs obsolete. We aren't there yet. Maybe we should start by applying this progress to software that already pretends to have AI, but in which the simple algorithms currently used are really not up to the task. It is hard to imagine a world in which AI is able to replace humans at doing real work, but not able to replace humans playing a computer game. If humans are content for MMO-like games, then partially replacing them by bots could significantly slow down negative network effects, and result in games that fail less quickly. That should be a worth while investment for game development studios.

The whole concept of "AI becoming unbeatable" relies on the assumption that "unbeatable", or maximizing win rate, is the only goal that can be set. But smarter goals could overcome that.

For example, for FPS games, you could have multiple models, "Beginner: Target 30% win rate" to "Pro: Target 75% win rate." The key difference is that the win rate is not a minimum but a target; it forces the models to keep in fallibility.

Strategy games are even better; using Civ VI as an example, you could specify "Science victory by turn 300; Domination by turn 400."

MMOs are more interesting - is there a goal for each of the player archetypes (explorers, min-maxers, etc.)? And how is that expressed vis-à-vis the different class types? It's a much harder problem and goal set to define.

But it's a good problem to go after. In the FPS example, I mentioned fallibility, which can be as simplistic as missing shots. But foibles, like healers over-pocketing the main tank, are much more interesting than simple fallibility. That creates a much more engaging game that would have better longetivity.
Post a Comment

<< Home
Newer›  ‹Older

  Powered by Blogger   Free Page Rank Tool