The final AI roundtable on Friday was AI for Beginners – we were in one room again, but it wasn’t as crowded. Some varied questions asked (not all of them for beginners, but a lot was explained for those that were). Dave Mark put up notes and audio from this roundtable too, if you want a better set of notes 😉

What do people think of Chris Heckler’s ideas?

• For graphics, we understand how to model worlds, with triangles and texture maps
• He suggests manipulating the AI in a photoshop like way
  • Neil suggests we won’t get there yet – not quite feasable yet if it is the right way to go

On nav meshes, changing a nav mesh with destructable terrain – it’s costly, how do we do it?

• Something in the Wisdom 4 books
• Middleware is working on it
• Some other companies are working on it
• Changable navigation meshes have been around at least a few years
  • Seeing a player jump through some points
• Company of Heroes – hierarchical pathfinding – Chris Journey
• Changing the nav mesh is dangerous – changing a key point means a lot of nodes on the mesh being made invalid
• Envelopes around destructible bits can help (Monster fighting through tokyo-like game).
• Making the nodes for points infinitely weighted for barriers
• Navmesh can have data added such as speeds of … (Not sure why I missed the end of this sentence…)

Will you be able to precompile some areas you know will be where the destructible terrain? Presetup.

• Yes. Method of efficiency

When saving the game, do you recompute or save the nav mesh?

• Should save the nav mesh saved – need to keep the state the same!

Anyone using middleware?

• Some using kygon for pathfinding
• No standard of AI like texture maps, and using middleware can lead to having to make a game to the middleware’s specs
• For most games, it’s hard to find a real middleware solution since you need to tweak the AI and so might as well build it from scratch
• Good to use middleware if you want to work on behaviours
• Resistance to neural nets in industry, but it will be used in some military simulator.
  • Game industry thinks the NN is the easiest technology to misapply.
  • Aversion to it since they think it’s awesome, fail at doing it, then don’t want to go back to it

What success stories there have been using NN?

• DoD, intelligence community – billions of dollars on applications of neural nets. A few isolated cases are productive, others failed.
• Neural Nets for wii motion detection
  • After filtering, it is 95% – needs to be 100% – trying to get there before shipping
• It is improbable to get 100% – the accuracy curve is a logathithmic scale in pain, and in user happyness.
  • This is for speech, but is likely the same for motion control
• PHD in neural nets – who shipped Colin McCrea 2.0 with neural nets, but he suggested nearest neighbour
• Instead of plain recognition, perhaps differentiation
• Applying neural nets need a lot of tweaking. Perhaps use it for what it is good at – representing interesting human action and reaction
• What are we trying to accomplish with neural nets? The right solution. B&W used reinforcement learning, not neural nets. More predictable and cheaper technology.
• Forza series – PHD in neural networks, learns how you drive your car.
• Tools needed, neural nets and algorithms are useful – not just one thing, but 75 for halo!
• Exhaustive training for it!
• Appropriate usage of it – a script helps in a rare situation for instance, but doesn’t scale
• Pre-ship or in game? Fine tune when testing/creation, but not after release.
• Will the tool save time? A tester might find the best track path faster going back to the racing car example.

Anyone doing speech recognition?

• Recognition mountain, then the natural language mountain is even bigger after it!

Smoothing AI outputs?

• Bioshock; animation system then locomotion system on top, so need to have a single action chosen not lots of them
• Subsumption will work with that
  • Or don’t be a dumbass rules!
• An AI wisdom article has something on applying Brook’s original subsumption to games.
• Need to add momentum to an action – “is it important for me to change what I’m doing?”

Anyone working with a memory constrained device?

• Not really. (The original asker is working on the DS)
• The question asker is working on the lower limits of enemies, etc.
• But look at old PC gaming articles – is it similar to a P90? P60? etc.
  • This keeps on recurring!
• Some articles in Wisdom books – such as not using A*
• Cell phone game used A* so it can work
• Constrained memory, CPU, so need to have the small 2-10% amount of CPU break up the work over frames (which is also thread friendly)
• Look for non-PHD implementations of mixing and matching the AI methods
• Or even fake it! – can you? or should you?

IntroGameDev.com (and AIWisdom.com) for article searches for older things.

Anyone used GPU for AI work?

• A single wisdom article on A* on it
• Tough since it’s all being utilized
• There is an instance that was explained – vector math is fast on it.

Academia and AI?

• EA in Alberta working on checking the way players played
  • Failure: a goal, so go back, find what went wrong so the AI can be fixed to not fail the time
  • NHL 08 shipped with it – possibly not the flow fields.
• University of Tokyo – worked with creation of characters.

How many companies are interested in an AI specializing programmer?

• Yes, lots!

And so ended the AI roundtable 🙂 Dave Mark got a recording, but it was a lot better structured (both my notes and the discussion) so it is easier to follow this time.