Ethical Decision Making Through Social Choice and Machine Learning
Abstract
As AI and society become more entangled, legitimate questions are being raised regarding the broader responsibilities of AI researchers and developers. The key question is this: How can we design ethical AI systems, which are aligned with human values?A possible answer is given by a new approach, which builds on an ancient concept democracy. Indeed, one of the basic ideas underlying democracy is that complicated, value-laden decisions can be made by asking a group of people to vote on the available alternatives. Consequently, researchers in computational social choice have argued that the field may provide tools for the design of ethical AI.In the last year, the PI has been developing and advocating the virtual democracy paradigm, which is an especially promising instantiation of this high-level idea. The approach consists of three steps: first, collecting preferences from voters on example dilemmas; second, using the data to learn preferencemodels, which generalize to any (previously unseen) dilemma; and third, at runtime, using the preference models to predict the voters rankings over the alternatives at hand, and aggregating the predicted rankings to reach a decision. In essence, the resulting systems hold a virtual referendum among modelsof voters for each decision they make.The PI and his collaborators have made initial progress on constructing the theoretical foundations of virtual democracy, and have built systems that apply the approach in two different domains: autonomous vehicles and food allocation. But the process of establishing virtual democracy as an effective andethical AI paradigm is still in its infancy; the proposal aims to take the next steps.The first thrust of the proposed research, explainability, will transform the way virtual democracy systems interact with humans. Indeed, one of the key advantages of the virtual democracy paradigm is that voting is a widely understood mechanism for decision making, opening the door to effective explanations. The proposed research will unlock the paradigm s explanatory power through solutions to algorithmic challenges.The second thrust, policy aggregation, aims to extend virtual democracy s reach to completely new domains. Specifically, a prominent approach to the design of ethical AI attempts to align the agent with human values by observing a person act in an environment, and using inverse reinforcement learning to derive her reward function and policy. Instead, the proposed research will align the agent with societal values by aggregating the policies obtained from multiple people.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012488
Entities
People
- Ariel Procaccia
Organizations
- Office of Naval Research
- President and Fellows of Harvard College
- United States Navy