The Mario Lives! video (and its follow-up video, Mario Becomes Social!) showcases an AI unit that is able to simulate emotional desicion-making within a virtual world, and can enter into "emotional states" such as curiosity, hunger, happiness, and fear. While this seems cool and exciting (especially for video game AI), I am confused how this would be useful in real-world scenarios.

What would be the point of building autonomous actors that would behave based on these emotional states, instead of simply knowing what they should do (either by hardcoding in the rules, or learning the rules through machine learning)?

  • 1
    $\begingroup$ Unless you provide some more information on the actual mechanisms used to model emotion, it's hard to say how this actually differs from the learned or hard-coded approach: for example, it could be that, despite the presence of the evocative labels 'happiness' and 'fear', the approach is essentially hardcoded. $\endgroup$ – NietzscheanAI Aug 15 '16 at 8:04

Humans have poor understanding of emotional rules. Probably every poster on here has experienced greatly misreading another individual emotionally. Further, people often don't act emotionally how they would expect themselves to act, for example we have all experienced frustration at someone else's irrational concerns and yet we are all guilty of holding irrational concerns of our own. This is the crux of why hard-coded emotional rules do not work - we do not have an understanding of what emotional 'rules' make someone feel real.

By moving towards autonomous state-based actors we move away from this issue. The actor's state transitions will of course be defined rules (hard-coded or learnt) but by abstracting the actor's emotional state from the specific context (e.g. specific actions trigger emotional state transitions which trigger responses, instead of a direct response to a specific action), the programmer prevents them-self from projecting their own beliefs/emotions/logic onto the actor.

Further, autonomous actors are more extendable. Consider an autonomous actor that is hard-coded to move towards 'upset' and 'angry' emotional states when experiencing 'pain'. Simply by associating a new world action with 'pain', one can trigger an emotional response from an autonomous actor that has not experienced that action before. When working in hard-coded emotional rules, this would not be possible.


Not the answer you're looking for? Browse other questions tagged or ask your own question.