At a recent AIMIA lunch for the Australian games industry, John Simpson had a chance to chat with Andrew Heath, Chief Operations Officer with the Melbourne-based developer Blue Tongue. Here’s what Andrew had to share. . .
Q: Do you have any current PC or console games in production that use AI?A: Jurassic Park: Operation Genesis (launch date Q1, 2003) is the main title we are working on at the moment, and certainly, it has some very advanced AI features. We're currently working on some other projects as well, but as yet we can't be more specific, but they also use some quite advanced AI.
Q: How is the AI in JP different to what we've seen previously? Have there been any major breakthroughs?A: The JP AI is some of the most advanced game AI technology seen to date. With this game we have based all AI of the park visitors and the dinosaurs on a complex set of heuristics, coupled with advanced Neural net programming -- all the behaviours of objects in the park are based on both 'natural tendencies' and on learned behaviour. Herbivores will naturally have a tendency to run away from carnivores, however learning that there is safety in numbers, flocking behaviour and defence is actually learned by the game objects. Also, if you build a park that is biased towards high level dinosaur action, you'll attract the thrill seekers to your park. If your park has an ecological bent, then you'll attract those visitors prefer the more natural attractions (read: greenies). None of the behaviours and interactions are pre-determined -- they occur naturally in real time.
Q: How do you begin engineering the AI? What steps do you take, from deciding how characters should behave, to actually seeing their behaviour on-screen?A: The Blue Tongue AI system was designed to meet a number of objectives:1. Decoupled interaction -- allowing an evolutionary approach to game balancing and design change, as new objects can be dropped into the world and are immediately recognised by other objects without any change to the existing object definitions.2. Data driven -- all objects are self-contained, and all balancing attributes are dynamically bound, substantially reducing the time taken for balancing iterations, and removing this responsibility away from coders entirely.3. Predictably unpredictable -- We want objects to behave in predictable ways, but want to also constrain this behaviour to designed limits to reduce the error space. A truly 'emergent behaviour' environment, with unpredictable results, is a testing nightmare. As emergent behaviour is by definition not dictated by design, controlling the results is typically intractable. However, there is an appeal in having interactions that, while bounded in possible outcomes, have outcomes that isn't. This helps create variety for the user.4. Constrained learning -- once again, the results must be constrained unpredictable. Learning is therefore limited to adaptation of perception. For instance, in Jurassic Park, dinosaurs become more wary of safari jeeps and ranger helicopters if they are hurt by them. Also important to this model is the notion of forgetting. This is another way of limiting scope for error.Based on these objectives, Blue Tongue created it's AI's Behavioural Drive System as part of its ToshiR engine. The basic AI flow is as follows:1. Establish the world2. Perceive the world3. Calculate Drive Magnitudes4. Determine the Primary Drive5. Execute the associated behaviour.Drives are the heuristics that determine the behaviour modes. These are things like Hunger, Thirst, Tiredness, Fear, Territory Protection, and Playfulness.
Q: Are there generic AI engines that are reused (eg. Doom 2 engine)? If so, which is currently the most capable, and why?A: There are a number of AI engines from middleware developers available, however at this stage most games have specific requirements of AI, so for some specific types of games, the might be useful, but in our experience there is none that compares to the capabilities of our senior programmers.
Q: How is the AI in computer games different to AI in robots (eg: AIBO)? How is it similar?A: The basic AI flow as above is pretty much the same whether you are developing games or programming robots. The main factors that affect the complexity are the number of drives, and the number of behaviours available to that which is being programmed. Obviously the more drives and the more possible actions, the more complex the AI matrix, and the behaviour learning matrix.
Q: There must be a great deal of research time involved with game AI -- where do you pull your research from?A: Very fortunately, Blue Tongue has staff that have had many years of experience in developing AI systems. At the moment, we do all our research and development on AI in-house.
Q: What are the major stumbling blocks in current AI technology? A: Really, the biggest stumbling block is being able to understand and keep track of the complexity of drives. By using Neural networks to calculate drives, obviously there ends up being a bit of a black box in terms of determining behaviours. At the end of the day, we are only limited by our creativeness, and the user's CPU.
Q: What do you see for the future of AI in gaming?A: The level of expectation of the public in creating realistic worlds is becoming higher and higher each day. I wouldn't say that there is necessarily any finite limit to AI, and considering that we only use about 10% of our own brain capacity, there is certainly the possibility of AI exceeding our capabilities.When it comes specifically to games though, predictability of behaviour can be a good thing, and it is all about the game experience. You still have to make the game fun. We think that we've got the mix just about right.
Note: The Turing Test is designed to determine if a computer program has intelligence. A person (A), a machine (B), and an interrogator (C) play an 'imitation game'. The interrogator can't see the other two, and tries to work out which one is the machine, and which is the human - only by asking them questions. If the machine can 'fool' the interrogator, it must be intelligent.
Issue: 137 | June, 2012