Robert Zubek chairs the panel on how we might next get to human-like AI. Behaviour is a big part of it, and I really enjoyed this talk. There’s discussion on the harder and easier ways to achieve things which is great.

• GDC Page

Notes

Robert – What interesting human level AI are you looking for? What’s the next big step?
Phil – What the player can perceive from AI. We all say we create deep AI but the player doesn’t see it. Research into if you can you do really subtle gestures if you do them consistently enough.
Richard – Maslows hierarchy of needs, motives form a pyramid – eating at the bottom, getting more sophisticated as it goes higher. Can get above fight and flee in characters.
Dave – When I’m playing a game I can see the rigid automatic behaviour then the transition to scripted sequences. Robot to hollywood actor. A consistent character in the gap.
Borut – Interesting to look at the Hierarchy of needs – I want to see the change in a character. Have some impact from the player change the character. Maybe as they learn meet more needs in the pyramid.
Dan – Feedback to the player, and how they receive that, is important. The other side is the figuring out of player intent that is needed to do.

Robert – Thinking of transitions and expressing behaviour to the player and higher level behaviours as well. What’s standing in the way?
All – Time.
Dave – In the middle of a pipeline – budgets on animation and other work – procedural animation and other things. Stuck in the middle of the way, how can you express it with in the budget?
Borut – Breaking out the pieces we can play with. A bottleneck with animation, but there is still things we can do and separate out from the content and should be exploring.
Phil – I’m exploring if you can be really iconic with the expressions, not ultra realistic but from incredibly simplistic images. Still some work to be done there to see how much simplicity you can get across to the player. A big variation there, and people will do different things in different games.
Richard – We use textual icons for The Sims, moodlets get shown. Can and do animated in text instead of waiting for the animations.
Dan – There’s a design constraint as well – the bastketball sport games.

Robert – A lot of animation goes towards the realism. Most of you have been talking about countering that…
Phil – Cartoons have been doing it really well for years, so if we don’t learn from Disney we are losing a lot of expression in games. It is all about expressing the mood, and to get more expression in scenes. More convincing and to care for them even if it isn’t realistic.
Dave – Shouldn’t talk ourselves out of things we want to do. The book Blink shows we take into account things subconsciously. A Japanese robotic person – manikin or person – actuators in the skin, low level, 70% went from manikin to 70% say human.
Borut – We need to however know how this works, even if the players don’t, just like how the lighting leads people.
Dan – If we’re going to do stylistic games need to get something on screen and iterate on top of since it will be subtle but needs checking.

Robert – A lot of this is on detecting higher level needs and broadcasting then seeing reactions. A lot of our AI doesn’t have this hierarchy of needs, do we need this?
Phil – We don’t have enough memory. Emotion doesn’t happen separate from memory, and you have to model the world around you through an emotional sense. It will constrain how much we can get into the belivability.
Borut – Difficult to pick a fiction that has NPC’s not dying quickly.
Dan – Hard to have a character on screen long enough. The player remembers along side the AI remembering. Need to have the AI try and think what the player thinks the AI remembers and do that.
Dave – Companions are by definition on screen all the time, and can add a level of depth when the player can change their treatment of them and they react.
Phil – An interesting thing Richard said, negative emotions in Sims 3 is brave. Negative character traits are people just going to get turned off by it? The idea of insanity in characters is interesting.
Dan – You can push characters but the designers and writers and animators need to be involved and fit in with the rest of the game. You can then grow the game significantly.
Richard – Sims has the 8 motives in Sims 2, they do in Sims 3 but in Sims 3 have different ones based on their personality traits. As well as each Sim having different motives to each Sim, they have different ones over time – short term and long term.
Borut – Do you do anything to balance the negative emotions out? Do they have any aims to do negative emotions?
Richard – No, haha.

Robert – Do you have any thoughts on the next level of low hanging fruit to get at? To get better, more humanlike AI?
Dave – Unpredictability. May have come from the QA or designers and be rigidly predictable. That gets the robotic feel. Humans are not completely rational – is it realistic to model a behaviour that is rigidly perfect every time? Is it okay to mix in some of that fuzzyness? We need to get away from rigidly predicable.
Robert – What about certain gams which needs designer puzzles or sequences?
Dave – Very Genre dependant, so it varies from game to game.
Dan – I’ve got a bit of a different answer. There’s a lot of stuff we can do but we have suspension disbelief from players. The focus of the universe if players – managing the game means looking at drama. Modelling players specifically – turtlers, people who cut you off in racing are good steps.
Phil – Giving designers tools for higher level goals and to give control, like a film director, then the AI sorts the low level out.
Richard – Having Facade they have provided a proof that there is an area you can move towards since it is possible.
Borut – The lowest hanging is bridging the scripted and normal behaviour.

Robert – What stuff really pays off – dependable techniques?
Dan – Barks are your best friends. Audio for specific moments really pays off.
Borut – Windows too, seeing the AI working.
Dave – Fuzzying stuff up on a bell curve in an easy way.
Phil – Make sure the characters facially track the things interesting to them, else they will look lifeless. Simple techniques – have some look away if shy, angry, doesn’t like the player etc.
Dan – Big bang from the book by tagging the world and having the people looking away when angry.
Richard – Putting effort into checking what needs to be look at changes.
Dave – Don’t need to necessarily look at the important things. Eye to eye contact is not constantly maintained in conversations in real life.