Unlocking the Potential of Global Human Expertise

Summary

This paper designs a framework called Realizing Human Expertise through AI (RHEA) to combine and distill solutions from a diverse set of models or solutions provided by experts. The steps are: Define Problem -> Gather Solutions -> Distill Model -> Evolve Solution. Examples were presented to demonstrate the framework's effectiveness in different domains.

Strengths

The author proposes a potential solution to collective knowledge that can be combined to discover solutions. It could be a way to find a better solution from the information stream when there is more than one solution.

Weaknesses

1. Some explanations of the framework are unclear and require additional details to better explain the "Distill" process. 2. The contribution paragraph needs more information to clearly describe the real contribution. 3. The conclusion is too broad and needs to be more specific.

Questions

1) The concern is the efficiency of this framework, as the Gather step requires waiting for experts to respond. The framework involves a human-in-the-loop process. How can bias be reduced when gathering solutions from the experts? (if these experts' solutions are biased or limited) 2) During a crisis, what happens if there are insufficient solutions? If there are two opposing solutions, what will be the outcome? 3) The description of the Distill process lacks sufficient detail for the reader to understand how it works.

Rating

5

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

This work has limited real-world evaluation. Fairness also needs to be considered in developing this framework.

Summary

The authors introduce a new evolutionary framework for combining expert policy predictions called RHEA and evaluate performance in both a synthetic domain that is highly interpretable and in a predictor from the XPRIZE Pandemic Response Challenge. The solutions are analyzed by qualitatively and quantitatively and compared against less sophisticated baselines.

Strengths

The paper is clearly written and focuses on an important problem. I liked that the authors explain the framework in the context of a highly interpretable example and then do significant qualitative and quantitative analysis on the solutions discovered by in the XPRIZE domain.

Weaknesses

See questions.

Questions

I did not get the sense that the “policy intervention problem” is something that is well studied. How is the problem statement here different from a contextual bandit or similar formalisms (knapsack problems?) Are there other baselines for this problem besides lesioned versions of the current framework? What are the computational differences between RHEA and the baselines was each algorithm run until convergence? Is it possible to show learning curves?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Summary

This paper introduces RHEA (Realizing Human Expertise through AI), a framework for combining diverse human expertise to solve complex problems using artificial intelligence. The key contributions are: * Recognizing the challenge of integrating diverse human expertise to solve global problems like those in public health. * Identifying requirements for an AI process to effectively combine human expertise. * Proposing the RHEA framework to meet these requirements, consisting of four steps: 1) Define the problem formally 2) Gather diverse expert solutions 3) Distill solutions into a canonical form (ie neural networks) 4) Evolve the distilled solutions to discover improved solutions * Demonstrating RHEA on a synthetic example. * Applying RHEA to a real-world problem: optimizing COVID-19 intervention policies. * Analyzing, for that real world problem: how RHEA recombines and innovates upon human expertise to discover improved solutions. The paper shows (at least for their examples) that RHEA can discover broader and more effective policy strategies than either AI or human experts alone. It highlights RHEA's ability to realize latent potential in diverse human expertise, even from solutions that may not seem immediately useful. The authors argue this approach could help bridge the gap between human-only and AI-only decision-making for complex global problems.

Strengths

Originality: The paper presents a novel framework, RHEA (Realizing Human Expertise through AI), which creatively combines human expertise with an evolutionary framework. This approach is original in several ways: The framework introduces a process of distilling human-created solutions into a canonical form (neural networks) that can then be evolved using population-based search. Quality: The authors provide a step-by-step explanation of the RHEA framework, though it’s unclear to me to how immediately useful those steps are as they are quite high level. * The framework is tested first on a synthetic example making it easier to understand. * The analysis of results on the COVID example is thorough, using multiple performance metrics and visualizations to support their claims. * The authors acknowledge some of the limitations and potential issues Clarity: The paper is generally well-structured and clearly presented: * The paper is well written. * The introduction effectively sets up the problem and the paper's contributions. * The illustration is helpful for explaining the concept. * The use of figures, particularly in explaining the RHEA framework and visualizing results, aids in understanding. Significance: The paper's significance lies in its potential impact on AI research and applications: * It addresses the important challenge of leveraging human expertise in AI systems, which is crucial for tackling complex global problems. * The RHEA framework could be applied to a wide range of domains beyond COVID-19 policy optimization. * The work demonstrates a practical approach to combining diverse expert knowledge, which could be valuable for collaborative problem-solving in various fields.

Weaknesses

Here's a substantive assessment of the weaknesses of the paper: * Limited generalizability: The paper's primary application focuses on COVID policy optimization, which, while relevant, may not sufficiently demonstrate the framework's broad applicability. The authors could strengthen their claims by: a) Providing a theoretical analysis for why RHEA should work in other domains. b) Discussing potential applications in other fields, with specific examples of how RHEA might be adapted. * Novelty: The evolutionary algorithm comes from other work so the idea is basically to initialize the parameters with distilled neural networks. This is a powerful way to solve this specific problem but given the limited analysis of the algorithm both empirically (when does it work) and theoretically, I am not sure if this work should be published at an ML conference or if it would be better at a more general purpose venue. * Insufficient comparison to existing methods: While the paper compares RHEA to evolution alone and distilled models, it lacks comparison to other state-of-the-art methods in policy optimization or ensemble learning. The authors should: a) Include comparisons to relevant baseline methods (e.g., reinforcement learning approaches, other multi-objective optimization techniques). b) Discuss how RHEA compares to other methods of combining expert knowledge, such as boosting or stacking. * Generalizability concerns: The paper does not sufficiently address what the operating regime is of RHEA. What problems does it work? What's the scale at which it works? Does it work if you have thousands of experts? * Lack of Theoretical guarantees: While the empirical results are promising, the paper lacks theoretical analysis of RHEA's properties. The authors could strengthen the paper by: a) Providing theoretical bounds on the performance of RHEA under certain conditions. b) Analyzing the convergence properties of the evolutionary process. c) Discussing how the distillation process affects the solution space and optimization landscape. * The structure of the paper: Some important details are in the appendix while there is a lot of analysis of the specific solutions and details (like the carbon footprint) which should be in the appendix. For example, as the specific ways this work is technically novel are unclear, the authors should move some of the related work to the main text. Addressing these weaknesses would significantly strengthen the paper, providing a more comprehensive and robust presentation of the RHEA framework and its potential impact. Post rebuttal update:

Questions

* How confident are you that RHEA can be applied effectively to domains beyond COVID-19 policy optimization? Could you provide examples of other complex problems where you believe RHEA would be particularly effective, and explain why? And also where it would not be effective? * Have you conducted any comparisons between RHEA and other state-of-the-art methods in policy optimization or ensemble learning? If so, what were the results? If not, which methods do you think would be most relevant for comparison? * How does the computational complexity of RHEA scale with the number of experts and the complexity of the problem space? Are there any foreseeable bottlenecks in applying RHEA to much larger or more complex problems? * Have you conducted any theoretical analysis of RHEA's properties, such as convergence guarantees or performance bounds? If so, could you elaborate on these? If not, what kind of theoretical results do you think would be most valuable to pursue? * How interpretable are the solutions produced by RHEA compared to the original expert inputs? Can you provide a detailed case study of a specific evolved solution, explaining how it combines and improves upon the original expert knowledge? * Could you provide more details on the distillation process? How do you ensure that the neural network accurately captures the essence of the expert's solution, especially for complex or nuanced strategies? For example, you find a Spearman correlation of 0.7 which does not seem particularly high to me. Also, I assume this 0.7 on the test set? What is the spearman correlation on the training set? * Does RHEA have mechanisms to incorporate new expert knowledge or feedback over time? How adaptable is the framework to changing conditions or new information?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors discuss the potential for approximation errors when applying the policies to the real world. The paper also mentions the need for user studies to fully evaluate the real-world effectiveness of RHEA prescriptors. The authors recognize that their cost measure was uniform over interventions, which may not reflect real-world scenarios. I think this is generally a good set of limitations. Potential areas where the authors could improve: * The authors could explore more deeply the potential long-term societal impacts of widespread adoption of AI-assisted policy-making tools like RHEA. * While briefly mentioned, the paper could elaborate on how to ensure the RHEA process remains transparent and interpretable to non-expert stakeholders. * The limitations section could address how RHEA might handle rapidly changing conditions or unexpected scenarios not covered in the initial expert knowledge.