Developer Journal: AI

Seed: "The best AI is a fun AI"

Believable NPCs, who live their own lives and who can talk to any player about almost anything in the game world, is one of the major features in Seed. This has prompted Runestone Game Development to carefully consider the entire take on Artificial Intelligence in their design.

“The all important factor is not to make the most realistic AI, but the most fun AI for the players to interact with,” says lead programmer Nis Haller Baggesen

In fact the AI realism has been ‘downgraded’ compared to the model constructed during the initial development period. The more realistic AI turned out to be harder to change according to story development and player interactions.

”The more realistic the AI, the more tools the writers need, and the more factors they have to take into consideration when writing stories and scripting NPC behaviour. The relatively complex language, we aimed for in the beginning, called for more AI interpretation and therefore more autonomous behaviour, again making it more difficult for the writers to control the reactions of the NPCs in various situations.”

“Actually it has been quite a challenge to construct a tool for letting the writers design NPC behaviour. AI techniques are not simple, and often include autonomous and random elements, not to mention a lot of tweakable parameters. Eventually, we came up with a solution where the writers obtain a large amount of control over AI expressions, which hopefully gives the NPCs more personality.”

The AI has been thought into the game from the very beginning. Along with CEO Lars Kroll, Nis Haller Baggesen holds a master’s degree in computer science and AI development.

“In my university projects the purity of the model has often been more important than the results. Or rather the objective was often to test a particular model rather than to get the best results. In games this is quite different. The result - the players’ perception of the AI - is all that is important,” Nis Haller Baggesen says.

He further endorses letting the AI ‘cheat’ – i.e. making use of knowledge, it might not be supposed to have.

“The AI in Seed may inspect a player’s inventory or a player’s relations, it may drastically change the mood of a NPC or create items for NPCs out of thin air – all in order to support a story.”

“Of course players sometimes complain about AI cheating in various games. However some of the things that NPCs do that are considered cheating (good aiming, fast control of multiple units etc.) would simply be considered high skill in a human player. In Seed 'cheating' will probably be even more acceptable, as the players are seldom in direct competition with the NPCs.”

One of the greatest challenges has been finding the right balance between scripted and autonomous behaviour.

“We want NPCs which are interesting and active, even when they are not currently engaged in a story, but on the other hand the writers need to have enough control over NPCs, when they actually take part in stories.”

Even though the NPCs have been designed to be as believable as possible, it is the players who are the main characters.

“Perhaps the most important task for the NPCs is bringing players together. Although it will be possible for a player to meet and interact with NPCs in every gameplay aspect, we believe the game to be most fun, when you are involved in stories with other players. This is also meant to boost the strong community feel we hope to achieve in the game,” Nis Haller Baggesen says.

He hopes for the coming Seed players to enjoy the NPCs.

“I don't expect the players to be blown away at how clever our AIs are. I do however hope that they will enjoy the freedom that the communication framework gives them, and that they will be positively surprised by how interactive the NPCs are. When they wander around and comment on their surroundings you can actually interrupt them and bring on just about any subject.”

- Nis Haller Baggesen is the lead programmer on the Seed MMORPG. He holds a PhD. part A (MSc equivalent) in Computer Science from the University of Aarhus. In 2001 he won the Learning Machine Challenge, an international contest for creating generic gameplaying AIs.

You can comment on this article here.