How to make AI perform well in games?

Good AI can add depth, challenge, humor and agency and there are some example of good AI such as thief, where AI actors respond in a realistic way to light and sound. Bad AI will probably ruin flow and immersion. There are not necessarily wrong but they break the core part of the game and make game experience less engaging. When designing AI, we should care about how the player perceives it.

Before designing AI, we should know some knowledge of our brains, they are incredibly stubborn which means they are not good at fighting natural tendencies. For example, it is hard to break a habit. Brains are hardwired to find patterns. Stereotypes allow us to efficiently navigate our world to deal with the onslaught of information. The smallest hint of repetition can be detected by the brain. Things stand out when they don't fit, because our brains have giant pattern matching systems which means they will something doesn't fit quickly. We always anthropomorphize. Brains attribute human-like behaviors and intelligent behaviors. And our memory schema is to weave narratives together.

We need to avoid the danger of uncanny valley concept. If we try to make something looks quite lifelike but not lifelike enough, brains are starting to put expectations on how real that should go but that can not fulfill the expectations. Then players will feel quite weird and creepy and try to avoid that. She gives a suggestion that behaviors are much better than a complex model and people will assume the complexity is there. So she suggests that do less and get more, which means start simple and resist complicating.

It is efficient to take 3 steps to design awesome AI. Firstly watch people play and try to know how they play and how they might expect other human to play. Secondly start stupid and give the brain the best chance to weave its own narrative. Keep iterating and try to identify appropriate and inappropriate behaviors and fix. One important thing is intelligent randomness, human brains are very sensitive to patterns and natural randomness is important for players to feel real.

In NBA Live, the opponent players are controlled by decent AI and you feel like competing with another guy. Your opponent will make decision of shooting, passing, blocking and assisting like a real human does. It is a good example of using simple appropriate behaviors and let players to weave their own narrative and naturally assume the complexity of the AI system. Each time you shoot or try to block, there is some randomness in your results which are similar to what would happen in real basketball match.

In order to improve current AI design, we could do more play test between two real players to find out their behaviors and when they behave that way. Comparing the matches with AI and with real human, we can find certain inappropriate actions that could be removed.