Image of a Terra Mystica board from Boardgamegeek.com

Learning from Weak AI

Unlike players, AIs don't have to internalize rules. They are the rules.

When it comes to tabletop board gaming, I currently have a crush on Jens Drögemüller’s and Helge Ostertag’s Terra Mystica. Terra Mystica is a strange creature of sorts as far as the strict definition of a Eurogaming goes. In board gaming circles European board games (as opposed to American board, also known to hardcore board gamers as Ameritrash games) are known for being light on thematic and narrative elements and heavier and more focused on the systems and mechanics that define the game’s play.

In regard to many of its systems, Terra Mystica does seem like a Eurogame. However, it has a fantasy theme (highly unusual for a Euro) and it has players take on the roles of different fantasy races, each with different abilities, strengths, and weaknesses. Such potential for lack of balance is more common to often more thematically oriented American games in which playing out a role is often more important than absolute fairness and balance.

Terra Mystica borrows from both traditions. It is kind of a civilization style game, which focuses on building and developing resources and an economy while jockeying for board position and territorial control. It’s quite good, and especially replayable because of the variety of strategies needed to learn if you are interested in playing a lot of different races, and, thus, some diverse play styles.

I’ve played dozens of games of the original version of Terra Mystica with a handful of people, and I just recently picked up the game’s expansion Fire and Ice, which features a new board and new races with new skill sets to master. So, I have some more learning to do with the game.

In my interest in doing so, I began looking online for a digital version of the game to mess around with. My gaming group doesn’t meet until the end of the week, and I just wanted to get a handle on some of the ways that the new races worked before explaining the new rules to my group. Since it was about three in the morning, while I did find an online multiplayer version used (among other things) for testing by the game’s developers, no one was online at the moment to play with.

So alternatively, I found a version of the game that I could play against AI, which seemed my only late night alternative.

Now, building AI for board games like Terra Mystica can be tricky. While Chess takes a lifetime for a player to master because it essentially boils down to a knowledge of move sets, AI can be developed that is pretty good at playing the game. With games like Terra Mystica, which are a blending of systems (area control, building an economic engine, resource management, etc.) along with some variable means of scoring points, which can change from game to game, building an AI that can “think on its feet” can be very, very hard.

And I discovered that the AIs that I was playing against were, well, just not very good at the game. I won every match that I played, even when purposefully playing races that are known to be extremely underpowered.

However, winning really was never my goal in my online foray. I came to learn, to learn a bit about some new rules that I had read and wanted to better understand in action, but I was also just curious to see how the computer would decide to prioritize its play in Teraa Mystica. After all, I’ve played a lot of games by largely learning alongside about a half dozen other new players. I wanted to see how our sense of strategies might or might not stack up against an AI developer’s sense of what should be prioritized in the game.

The AI, of course, couldn’t win against me, but I could see why it had a problem in doing so. It clearly lacks the forethought to plan for the contingencies in a game where a whole lot of systems can create a lot of variable possibilities and then to decide how to evaluate which systems take priority over others under very specific circumstances. AI can be good at analysis, but humans tend to be better at evaluating or making judgments about values better than those pesky machines do. That being said, I got kind of fascinated watching the way the AI opened the game, and how that differed a bit from how me and my friends have approached our own games. It very clearly made mistakes later in the game, but I suspect that some of the variable openings that I watched probably could lead to some approaches to play that I hadn’t considered. It’s analysis of the overall game was probably good. It just couldn’t follow through when things got messy.

Oh, I also learned that I had misunderstood two rules about the game, simply by playing against the AI and discovering that I couldn’t do a couple things that I had thought were legal moves. One of those rules probably hasn’t effected many the outcomes of my previous games too much, but one of them will be a bit of a game changer when I have to make the correction the next time I play with my gaming group.

And these are the moments that make me value playing against an AI, even a weak one, not because I necessarily feel that I will be especially challenged by them, but because they make me reflect on my own analysis of a set of rules. Human beings learn rules and then have to internalize them. That internalization improves play, but also requires a level of subjective interpretation and evaluation that can lead to misapprehension of the intent of those rules. It can also lead to an insistence on approaching play in ways that seem to work best to us, as framed by our own personal internalization of them and our own insular position within a smaller group of players interpretations of a game and how it should best be played.

AIs don’t have to internalize rules. They are the rules. The rules are already present within them. They define their “identities” as psuedo-intelligences and determine their behavior as a result. There is no way that they can evaluate them, or at least not as the player is capable of doing.

So, I return this weekend to my human opponents after a string of fairly easy victories in a game that I usually find much more challenging and interesting. However, I do have the ability to evaluate weakness, to shift from an incorrect rule set to another, and even to consider how a losing opening played out by a machine might lead to an improvement in my own play. Let the computer analyze, and I’ll be quite happy to find new ways to utilize the data by changing my own relationship to the rules that govern me.