Dealing with Imperfect Rationality in AI Systems

AI systems often collect their input from humans. For example, parents are asked to input their preferences over primary schools before a centralised algorithm allocates children to schools. Should the AI trust the input provided by parents who may try to game the system? Should the parents trust that the AI system has optimised their interests? Would it be safe to run the algorithm with a potentially misleading input?

Algorithmic Game Theory (AGT) is a research field that attempts to add safety and trustworthiness to AI systems vis-a-vis strategic reasoning. With its set of symbolic tools, one aims to align the goals of the AI system (e.g., the allocation algorithm above) with those of the agents (e.g., the parents above) involved. The AI will then be safe, in that we can analytically predict end states of the system, and trustworthy, since no rational agent will attempt to misguide the system and the system will work on truthful inputs.

One assumption underlying much of the work in AGT is, however, pretty limiting: agents need to be fully rational. This is unrealistic in many real-life scenarios; we, in fact, have empirical evidence that people often misunderstand the incentives and try to game the system even when it is against their own interest. Moreover, modern software agents, often built on top of AI tools, are seldom able to perfectly optimise their rewards.

This project will look at novel approaches to deal with imperfect rationality, including analysis of known AI systems and the design of novel ones. This will involve theoretical work that builds on the recent advances on mechanism design for imperfectly rational agents (namely obvious strategyproofness and not obvious manipulability) to include more complex domains and the modelling of further behavioural biases in mechanism design.

https://kclpure.kcl.ac.uk/portal/files/174629484/OSP_MOR.pdf

https://link.springer.com/article/10.1007/s00224-022-10071-2

https://arxiv.org/pdf/2202.06660.pdf

https://arxiv.org/pdf/2007.11868.pdf

Project ID

STAI-CDT-2023-KCL-13

Supervisor

Prof Carmine Ventrehttps://www.kcl.ac.uk/people/carmine-ventre

Category

Reasoning