David Roberts, Charles Isbell, Mark Riedl from the School of Interactive Computing, Georgia Institute of Technology.

As a field, artificial intelligence (AI) has been applied to games for more than 50 years, beginning with traditional two-player adversarial games like tic-tac-toe and chess and extending to modern strategy games, first-person shooters

I should finish this blockquote off (and add it for the others). Opinions on this: Interesting basis of having an interesting adversarial game (not just one the AI must lose – thus not intentional artificial stupidity) which might be applicable to RTS games more then anything else.

Content

Looking at game AI as a storytelling title – a little afraid of putting forward “Why a game of chess is a story”, but that is the example game used.

Currently AI is concentrating on adversarial games – the AI is trying to beat the player, it’s win or lose. Recent work on behaviours of adversaries. This doesn’t apply to this though, don’t want a harder game want a better story in the game.

Chess has a total amount of states numbering likely more then the amount of atoms in the universe, so computers can’t perfectly compute a game. Use heuristics instead to perform decision making. Minimax search example, with a 2×2 Tic-Tac-Toe (where you must get two X or O next to each other) expands the entire search space until the games end. Since the first player always wins if played perfectly optimally, it is not fun for the player – so need to add intelligent mistakes or artificial stupidity. A well fought battle victory is better then an easy win – fun vs. hard.

Can do a variety of behaviours to implement mistakes. Missing the first shot, aborting last minute shooting in FPS’s, AI that knows the best path and intentionally doesn’t choose it, and other things.

Optimal now is defined as the quality of experience of the player, to make interesting games for the player. Makes trajectories of a game.

For instance, a mystery novel – starting out with a small unrevealed event, with additional reveals and red herrings – then the denouement to bring down the tension. If not managed for a interactive experience then this can radically change (see changed graph) where it might be more appropriate for a horror film or something.

For this AI and storytelling, need to keep coherence between events, taking some parts of narrative design and apply it to adversarial AI like Chess. There are a lot of trajectories of chess, so which are good and bad? Need to inspect specific parts of the trajectory and see specific coherent attacks and moves in a certain area.

Chess players themselves describe a game with narrative story elements – forks, pins, discovered attacks. Example chess board – white has moved his knight forward, black can move his pawn forward to attack the knight which means that will be taken. If the player is white, good for the AI to recognise, but if player is black, the AI must realise it can happen and not do it.

“Intelligent mistakes” doesn’t necessarily apply since it is more how the story progresses not the amount of wins/losses. Chess has a lot of structure so attractive to AI developers, but can apply it to RTS games. Like chess, can use heuristics to represent the game that enables the same basis of trajectories and apply the same type of narrative reasoning.

For instance, the number of units over time, and amount of resources available – can have graphs showing how well the player’s graphs being inspected over time and having the AI make changes of how it goes.

Questions

Is there any way of adding dynamic situations instead of static sets to try and achieve?

Nothing explicitly – plan recognition that is one area that might be applicable.

How do you choose the way to make decisions for the AI?

It depends on the game design,

More fun for whom? Different players have different experiences of fun.

Can tweak it and add user modelling into it, not explicitly in the paper.

From design, how about the game becoming so dumb if the situation is changed and much easier so the player doesn’t improve?

Repeated episodes is a whole another thing to think about to get the goals over time changed.

How do you deal with the player radically changing from the expected graphs?

There is work on steering players towards a path, and in narrative design there is contingency planning. It is slightly less obvious in an adversarial setting like chess.