Prospect theory and utility functions

From Red Blob Games's Blog

(Note: I'm going to try writing more unpolished things on this blog and leave the polished articles to redblobgames.com.)

The utility functions I learned in economics are history-agnostic. They look at the current state of the world and calculate a “utility”. For example, you might say the utility of money is the logarithm of the amount of money you have.

Prospect Theory says that this view of the world does not match how people actually behave. Instead, the history of how you got to a certain point matters in how you value it. There's an asymmetry around a “reference point”:

(credit: Wikipedia, Prospect Theory page)

Consider these scenarios:

• You get $200 and have a 90% chance of losing $100 of it.
• You get $100 and have a 10% chance of gaining an additional $100.

These are mathematically equivalent. Both have a 90% chance of giving $100 and 10% chance of giving $200. However, they are not equivalent to humans. That's because humans consider not only the final result but how it was reached. Having $200 and then losing $100 feels different from having just the $100 in the first place. Even though the outcomes are the same, the reference point is different, so it feels different. Prospect theory takes this into account somewhat.

In game AI, I've only used regular utility functions. However, it seems reasonable to try using prospect theory in some way. Even prospect theory isn't complete; there are more human behaviors in decision making and valuation that it doesn't account for. Maybe FTT or something else. But sometimes you want to balance simplicity and comprehensiveness. In any case, it's something I'll want to ponder the next time I'm writing AI evaluation functions for NPCs.

Labels: design , ideas , math