Designer’s Diary: Can AI Beat Its Creator?

Probably no other phenomenon in the IT world generates as many articles as generative AI of all kinds. And our current Designer’s Diary from the development of Lost Ruins of Arnak will be no exception. This time, however, we might be talking about a slightly different kind of AI than what’s been trending in recent months. This artificial intelligence won’t chat with you or recommend what to visit during a long weekend in Prague. But it will play a balanced game of Arnak with you.

Although it may seem like AI has only been around for a short time, that’s certainly not true in the gaming industry. AI opponents—computer-controlled players—have been with us practically since the beginning of gaming. And Lost Ruins of Arnak is no different. If you want to explore the island of Arnak, but your friends don’t have time for a game or two, artificial intelligence can step in.

“Let's get one thing clear at the beginning: it’s not a neural network. Our AI is—let’s say—built in a traditional way. It doesn’t learn. All its strategies are already encoded within it. And thanks to its speed, it can simulate many moves ahead and evaluate which one is the most beneficial,” explains Matúš Kotry, one of CGE’s programmers responsible for the AI. He has already created a player-favorite AI for Through the Ages. Now he’s working on Arnak.

One Algorithm to Rule Them All

“The backbone of any such AI is a tree-search algorithm. In our case, it’s Monte Carlo Tree Search. At the beginning of your turn, you always have a wide range of options—playing a card, advancing on the research track, discovering a location, etc. The opponent can then react in various ways. And so on. All these possibilities of how the game can unfold can be imagined as a tree,” explains Matúš.

An average game of Arnak lasts about 40 turns, which in practice means there are roughly 20⁴⁰ possible ways the game might go through. That’s many orders of magnitude more than the number of stars in the observable universe.

The magic of creating a good AI lies in deciding which paths to prioritize and which to ignore. “Many of those moves simply don’t make sense at a given moment, so it’s not worth spending time analyzing them at all. So I basically give the AI strategic hints, some of my own know-how. But then it uses that knowledge to find the ideal move,” Kotry adds.

A good example is discarding cards: “If the AI has the option to exile a fear card, it will always do it. It doesn’t need to ‘think’ about it, because this decision is almost always better than exiling any other card in the game.”

How Many Moves Ahead?

So how many moves does the Arnak AI consider before making a decision?

“It varies. The number of evaluated moves is one way to adjust difficulty. At the highest difficulty, it considers about 10,000 possible game developments each turn,” says Kotry.

Each decision node in these simulations is evaluated based on the quality of the outcome. This allows the AI to identify truly strong moves and choose one. However, it doesn’t simulate the entire game to the very end. Its calculations stop at the end of the current round.

“I decided there’s no point in thinking further. At the end of the round, you draw five new cards, the card row refreshes, locations reopen—there are too many random elements.”

Memory of a Goldfish

Is it any problem for AI’s plans when a player disrupts it?

“Actually, that’s not a problem at all. It could be surprising, but in fact, the AI doesn’t remember its original plan. Whenever it gets its turn, it recalculates everything from scratch, regardless of its previous move,” explains Matúš.

And yet, it appears as if it has a long-term plan. That’s simply because, unless the situation changes, the recalculation will likely produce the same result as before—even if the AI itself doesn’t “know” it.

“I’d say it does have a plan. But unlike a human, it’s more flexible. When the opponent ruins its intention, there’s no table flip. It just recalculates and adapts instantly,” he says. This also allows the AI to take over a game at any point. For example, when a player drops out of an online match.

Master and his padawan

Can the AI beat its creator—or even come up with something the programmer didn’t anticipate?

It might seem unlikely. After all, the strategic hints and know-how come from the developer. But the AI’s ability to surprise lies in the huge number of situations and combinations it can calculate.

“As a player, I might think about three strategies. By the time I consider the third, I’ve already half-forgotten the first. That doesn’t happen to AI. Thanks to that, it can discover things I thought weren’t even possible,” admits Matúš.

“In one of my recent games, the AI managed to acquire both assistants in the first round. Anyone familiar with Arnak knows how hard it is to obtain assistants. I thought it was impossible. Then I found statistics on BoardGameGeek showing that top players manage it in about 3% of games. So I let AIs play hundreds of matches against each other. To my surprise, they were able to do that in… about 3% of cases. So, I ended up programming an AI that discovered a tactic I didn’t even know existed.”

The Search for One Extra Point

A large part of the challenge lies in balancing difficulty levels. The AI should be neither unbeatable nor too easy.

“Even on the easiest difficulty, the AI can’t play in an obviously stupid way. That wouldn’t be satisfying. For example, if it discarded its strongest card instead of a fear card, that would be a legal move—but it would just make players shake their heads,” Kotry explains.

His solution? “I try to make it avoid mistakes in short-term decisions but introduce imperfections in long-term ones. For example, I add a random factor to move evaluation, so it won’t always pick the absolute best move—just a slightly worse one. But only for main actions. Once it commits, it executes everything else optimally.”

Another way to adjust difficulty is to reduce the number of evaluated moves, increasing the chance it won’t find the best solution.

“I regularly test it by letting three AIs play against each other about 20 games, which takes them just a few minutes by the way. Then I compare their average score to data from real players,” he says. “The analysis based on 12,000 games published by Akrnicologist on BoardGameGeek helps me a lot with it. He made an article with a lot of useful data, which I’m using as a reference. Thanks to him, it’s much easier for me to compare my AI to the real players of the different skill levels.”

Currently, lower difficulties perform similarly to beginner and intermediate players, while the hard difficulty averages around 80 points—sometimes more.

“If the AI has a good game and a top player doesn’t, it could win. So it may not be at the absolute top level yet, but it’s definitely a worthy opponent for anyone,” says Kotry.

Souls-like difficulty, anyone?

Could the AI be made even stronger?

“Yes, but improving it by just one point on average is already quite hard. For example, doubling the number of analyzed moves from 10,000 to 20,000 only improves it by about one point,” he explains.

Still, Matúš has ideas for further improvements.

“If I manage to create an AI that averages around 87 points, I’d probably introduce it as a fourth difficulty level—with a warning that it’s brutally strong. Or maybe unlock it only after players repeatedly beat the hard difficulty. I want Arnak to be a challenge for everyone who dives into it.”