AI with conflicting objectives?

Question

A recent question on AI and acting recalled me to the idea that in drama, there are not only conflicting motives between agents (characters), but a character may themselves have objectives that are in conflict.

The result of this in performance is typically nuance, but also carries the benefit of combinatorial expansion, which supports greater novelty, and it occurs to me that this would be a factor in affective computing.

(The actress Eva Green is a good example, where her performances typically involve indicating two or more conflicting emotions at once.)

It occurs to me that this can even arise in the context of a formal game where achieving the most optimal outcome requires managing competing concerns.

• Is there literature or examples of AI with internal conflicting objectives?

Answer 1

There are multi-objective optimization problems, where the objective functions may be in conflict with each other, which can potentially have multiple Pareto-optimal solutions. The paper Multi-objective optimization using genetic algorithms: A tutorial (2006) gives a good overview of the multi-objective optimization problem with genetic algorithms, which can be called evolutionary multi-objective optimization (EMO) or multi-objective optimization evolutionary algorithms (MOEAs).

A common multi-objective genetic algorithm is NSGA (or NSGA-2 and NSGA-3), which stands for Non-dominated Sorting Genetic Algorithm, which is based on the concepts of non-dominated sorting, Pareto front and optimality, niches (sub-populations), and elitism (the best individuals of the current population are carried over to the next generation).

If you want to play with MOEAs, you may wanna try the Python deap package, which supports, for example, the NSGA-2 algorithm.

Answer 2

MOEAs sounds very cool, but I feel that you can't really talk about conflict in AI without discussing generative adversarial networks (GANs), which have been shown to have amazing performance by training a model to say detect in-between pictures of cats and dogs and an adversarial network being trained to create pictures to attempt to trick the training network as much as possible. The completely conflicting objectives of the networks enable both to be trained very well so the models, in the end, are much more robust and able to handle sometimes bizarrely generated edge cases.

I also found this paper Evolutionary Multi-Objective Optimization Driven by Generative Adversarial Networks (GANs) (2019), which combines MOEAs and GANs, but there are potentially more related papers.