Theory for advanced AI in games

fr0st2k · 2011-08-12T17:44:45

Most developers deal strictly with stats, and don't expand too far past that. Stats are vital, of course, but so is the layer on top of them; the personality. Psychology can help define how an AI interacts with the world around him. The personality should be very dynamic, and flow depending on the players actions, environment, or other AI. Personality should derive from what the developer feels are necessary decision making traits for their AI. If for example, we look at the game Oblivion, the AI was developed to walk around town and look busy. However, it was far too scripted, and the AI was about as basic as AI could get. Of course, processing power, and the like plays a role in these calculations, but for this example, lets ignore it. In order to spice it up, You can assign a few personality roles to each AI. For simplicity, lets just use 3. Friendly, Angry, Neutral. Then below that layer, add stats. The number of stats, like their personality traits, is limitless. The more you have, the more complex the AI becomes. Stats for each AI could include things like [player/environmental interaction stats]: bad outcome in last hour [negatively affects happiness] good outcome in last hour [positively affects happiness] interactions [if friendly, + affects happiness. if angry, - affects happiness] Distance traveled [boredom, fatigue] etc. Stats would be recorded for whatever the designer feels is important. Each stat will affect a preconceived list of "decision making" stats or "interaction stats. When an AI interacts with a player or something in the environment, these "interaction stats" would be used in the computations for success or failure. Back to the Oblivion example, The designer will originally denote a personality to a NPC. For this example, they choose "Friendly." Friendly's internal code will interact differently with the stats. If interactions is very high for the day, then a Friendly NPC will be happier. The code would look something like: FriendlyValue*Interactions = happiness (extremely simple, but you get the idea). Now you can hardcode quests/secrets into the AI for the player to interact with. If AIhappiness > 50 then respond with "Here take this free item"! While the outcome of this system is similar to what is typically done, the expansiveness is unparalleled. You can make an extremely complex AI using by: 1) Assigning a Personality Trait or traits 2) Assigning important stats to keep track of 3) developing "interaction" formulas that relate traits and stats to what the AI interacts with. With the basics lined up, I'll try and do a full blown example. I'll start by listing the AI traits, and the requirements of the object the AI will interact with. AI: -Trait Friendly - 80% Anger - 5% neutral - 15% Stats (based on last hour) -Drinks: 0 -Food: 5 -Interactions: 3 -failed interactions: 8 -money: 5 -good outcome: 0 -bad outcome: 8 -neutral outcome: 3 Decision makers (out of 100%): -Thirst - 20% -Curiosity - 60% -Boredom - 80% -happiness - 50% Thirst is derived from the algorithm: on update: (oldthirst+(( drinks*.02)-.05)) So, it will always decrease by 5% if the NPC has no drinks. Traits have no bearing on thirst, so its not factored in. AI gets buzzed at 20% thirst to find something to drink. So he jumps into the pathing routine and maps out his way to the closest vending machine. He walks towards it as the player watches..."i wonder where hes going?" The AI arrives. Now he interacts with the vending machine code: Vending machine code: -Refresh cans once a day Current Cans: Sprite:0 Coke:3 lipton:1 On proximity: Spark AI "am i thirsty?" routine. Essentially, when an AI passes in 'x' proximity of this object, it alerts it to check and see if its thirsty. This check can be complicated or simple, but will be based on the derived decision making stats of the AI. Obviously we'll want to check "thirst" decisioncheckvalue = 100-(.2[thirst value]*100) randomized decision = decisioncheckvalue*rnd(+-5%) [can base this on anything..including other stats...like "cheap" which would relate to dollars] if randomized decision > 70 buy drink the AI would have an 80 stat, plus or minus 5%, thus they would get a drink. However, our AI already is headed there for a purpose, and thus, would skip the close proximity reaction (which would occur for almost all objects in the game that the AI can interact with.) The AI arrives, and already wants to buy a drink. Maybe in his AI, its designated that his favorite drink is Sprite. Whenever he drinks spite, he gets a good outcome. Whenever he drinks anything else, its a bad outcome. So he buys sprite, but its empty. +1 bad outcome, then buys coke, another +1 bad outcome. The bad outcomes readjusts his happiness, and it lowers. Now a little kid comes by who is scripted to ask players/AI's for drinks. He hits the radius of the AI and buzzes that he has a drink, and starts the interaction. Just like the vending machine, there is a scripted "check" to see if the AI will give it to him. To make it as simple as possible, it just checks his "happiness" If 'friendly' then it checks if Happiness is > 50%. If yes, then give soda. Well, out AI's happiness was just adjusted due to 2 bad outcomes, so his happiness falls below 50% and he does not give the soda. The boy thus, gets a bad outcome. We can even change up personality traits. If Happiness < 20% for 2 hours, then change trait to angry. This is a hard concept to explain in a small forum window. Also, since I haven't developed an actual game yet, the examples are very loose. However, im sure that if you reread it a few times you can see where I am going with it. Essentially, in short: the idea is to develop AI with different layers that affect each other. Would love some discussion on the subject! Let me know what you think.

Artificial Intelligence Programming

Started by fr0st2k August 11, 2011 06:06 PM

11 comments, last by IADaveMark 13 years, 6 months ago

IADaveMark

3,839

August 12, 2011 12:46 AM

I believe Szecs was referring to the OPs other thread, specifically.

Yeah... hadn't read the other thread at that point. My bad.

Dave Mark - President and Lead Designer of Intrinsic Algorithm LLC
Professional consultant on game AI, mathematical modeling, simulation modeling
Co-founder and 10 year advisor of the GDC AI Summit
Author of the book, Behavioral Mathematics for Game AI
Blogs I write:
IA News - What's happening at IA | IA on AI - AI news and notes | Post-Play'em - Observations on AI of games I play

"Reducing the world to mathematical equations!"

Dave Weinstein

1,313

August 12, 2011 04:56 PM

If memory serves me, terrain based AI was in use in commercial games in the 1980s.

IADaveMark

3,839

August 12, 2011 05:44 PM

Forgot to add Creatures to the pile of games using "let's allow objects to tweak numbers so that the agents change behaviors" technique. And that was how many years ago?

Dave Mark - President and Lead Designer of Intrinsic Algorithm LLC
Professional consultant on game AI, mathematical modeling, simulation modeling
Co-founder and 10 year advisor of the GDC AI Summit
Author of the book, Behavioral Mathematics for Game AI
Blogs I write:
IA News - What's happening at IA | IA on AI - AI news and notes | Post-Play'em - Observations on AI of games I play

"Reducing the world to mathematical equations!"

Theory for advanced AI in games

This topic is closed to new replies.

Popular Topics

Recommended Tutorials

Theory for advanced AI in games

This topic is closed to new replies.

Popular Topics

Recommended Tutorials

Reticulating splines