Machine Learning | UCL UCL Department of Science, Technology, Engineering and Public Policy

Machine Learning for Unlocking the Policy Impact of Transdisciplinary Research

By Basil Mahfouz, on 8 July 2024

As the 31st International Conference on Transdisciplinary Engineering 2024 kicks off at UCL East, over 100 engineers from around the world—including the United Kingdom, Brazil, Mexico, United States, China, Japan, Sweden, Singapore, and others—converge in London to explore how transdisciplinary engineering can drive social change and improve the world. This conference provides a platform for discussing the crucial role of engineering and science in addressing societal challenges through innovative, interdisciplinary approaches.

We already know that transdisciplinary research teams comprising engineers and others, tend to produce research that is more likely to have policy and commercial impact. Yet, with over 334 categorized research fields, there are more than 35 billion possible combinations for interdisciplinary work for teams of up to 5 researchers. Depending on the combination of researcher capabilities, some interdisciplinary teams may be better suited for disruptive science, developing patents, or informing policy. But which combinations of fields lead to which type of impact?

Supported by Elsevier and working with the Growth Lab at Harvard Kennedy School, we are applying complexity methods and machine learning on bibliometric data to understand which combination of researcher capabilities leads to high-impact research. For this blog, we’ll discuss our work within the context of the impact of interdisciplinary climate research on public policy.

To calculate interdisciplinarity, we determine the capabilities of authors based on their publication history in different fields. Each author is represented by a vector indicating the number of times they have published in each field. These author vectors are then used to calculate the disciplinary diversity (DDA) of each paper, reflecting the combined expertise and capabilities of the co-authoring team.

As a first step, we ran a series of statistical analyses and regressions to evaluate the relationship between a paper’s interdisciplinarity score and the number of policy citations it received. Preliminary results show that interdisciplinarity explains almost 15% of the variance in policy citations, making it the strongest predictor of policy impact we have identified so far. In fact, transdisciplinarity is found to be three times better at predicting policy citations than conventional metrics of research excellence, exceeding the combined effects of academic citations, journal impact factor, and author h-index.

We then aggregated our analysis at the topic level by calculating the average interdisciplinarity score of papers within each topic. The first observation is that climate change-related topics with high interdisciplinarity are less common than those with low diversity. The distribution of these topics is shown in Figure 1.

Fig 1: Distribution of paper interdisciplinarity in climate research

The second observation is that topics with higher ratios of policy citations have nearly double the average interdisciplinarity score compared to topics with relatively low policy citations. Figure 2 illustrates the difference in interdisciplinary distribution between the low and high policy relevance groups.

Figure 2: Interdisciplinary Distribution by Policy Relevance

Finally, using the paper vectors, we are developing a machine learning model to understand which combinations of author capabilities and team dynamics lead to high policy impact. The preliminary model has already found that for climate change-related research overall, teams involving researchers with expertise in economics, meteorology & atmospheric sciences, general & internal medicine, ecology, and horticulture tend to be associated with higher policy impact.

Moving forward, we are refining the model to predict the optimal team dynamics for high impact within specific policy topics. We are expanding the training data to include additional metrics and features, such as academic seniority, international collaboration, and research text. Furthermore, we will incorporate the distances between research fields to study the effects of deep interdisciplinarity, bringing together researchers from relatively less connected fields, on social impact.

As TE2024 brings together engineers and researchers from around the globe, this work exemplifies the potential of interdisciplinary collaboration in addressing global challenges and driving social change through innovative, data-driven approaches. Equipped with machine learning tools like this, researchers can develop targeted strategies to form data-informed transdisciplinary teams, optimized for maximum societal benefit and impact.

Filed under Education, Policy Impact Unit, Research

Tags: Machine Learning, Policy Impact, research, transdisciplinary

No Comments »

Adversarial Attacks, Robustness and Generalization in Deep Reinforcement Learning

By Ezgi Korkmaz, on 20 December 2023

Reinforcement learning has achieved substantial progress on successfully completing tasks, from solving complex games to large language models (i.e. GPT-4) including many different fields from medical applications to self-driving vehicles and finance, by learning from raw high-dimensional data with the utilization of deep neural networks as function approximators.

The vulnerabilities of deep reinforcement learning policies against adversarial attacks have been demonstrated in prior studies [1,2,3,4]. However, a recent study takes these vulnerabilities one step further and introduces natural attacks (i.e. natural changes to the environment given that these changes are imperceptible) while providing a contradistinction between adversarial attacks and natural attacks. The instances of such changes include, but are not limited to creating a blur, introduction of compression artifacts, or perspective projection of the state observations at a level that humans cannot perceive the change.

Intriguingly, the results reported demonstrate that these natural attacks are at least equally, and often more imperceptible compared to adversarial attacks, while causing larger drop in policy performance. While these results carry significant concerns regarding artificial intelligence safety [5,6,7], they further raise questions on the model’s security. Note that the prior studies on adversarial attacks on deep reinforcement learning rely on the strong adversary assumption, in which the adversary has access to the policy’s perception system, training details of the policy (e.g. algorithm, neural network architecture, training dataset), and the ability to alter observations in real time with simultaneous modifications to the observation system of the policy with computationally demanding adversarial formulations. Thus, the fact that natural attacks described in [8] are black-box adversarial attacks, i.e. the adversary does not have access to the training details of the policy and the policy’s perception system to compute the adversarial perturbations, raises further questions on machine learning safety and responsible artificial intelligence.

Furthermore, the second part of the paper investigates the robustness of adversarially trained deep reinforcement learning policies (i.e. robust reinforcement learning) under natural attacks, and demonstrates that vanilla trained deep reinforcement learning policies are more robust than adversarially, i.e. robust, trained policies. While these results reveal further security concerns regarding the robust reinforcement learning algorithms, they further demonstrate that adversarially trained deep reinforcement learning policies cannot generalize at the same level as straightforward vanilla trained deep reinforcement learning algorithms.

This study overall, while providing a contradistinction between adversarial attacks and natural black-box attacks, further reveals the connection between generalization in reinforcement learning and the adversarial perspective.

Author’s Note: This blog post is based on the paper ‘Adversarial Robust Deep Reinforcement Learning Requires Redefining Robustness’ published in AAAI 2023.

Author Website: https://ezgikorkmaz.github.io/

Twitter: https://twitter.com/EzgiKorkmazAI
Paper Abstract: https://ojs.aaai.org/index.php/AAAI/article/view/26009
Paper Link: https://ojs.aaai.org/index.php/AAAI/article/view/26009/25781
Reading List on Adversarial Reinforcement Learning: https://github.com/EzgiKorkmaz/adversarial-reinforcement-learning

References:
[1] Adversarial Attacks on Neural Network Policies, ICLR 2017.
[2] Investigating Vulnerabilities of Deep Neural Policies. Conference on Uncertainty in Artificial Intelligence (UAI), PMLR 2021.
[3] Deep Reinforcement Learning Policies Learn Shared Adversarial Features Across MDPs. AAAI Conference on Artificial Intelligence, AAAI 2022. [Paper Link]
[4] Detecting Adversarial Directions in Deep Reinforcement Learning to Make Robust Decisions. International Conference on Machine Learning, ICML 2023. [Paper Link]
[5] New York Times. Global Leaders Warn A.I. Could Cause ‘Catastrophic’ Harm, November 2023.
[6] The Washington Post. 17 fatalities, 736 crashes: The shocking toll of Tesla’s Autopilot, June 2023.
[7] The Guardian. UK, US, EU and China sign declaration of AI’s ‘catastrophic’ danger, November 2023.
[8] Adversarial Robust Deep Reinforcement Learning Requires Redefining Robustness, AAAI Conference on Artificial Intelligence, AAAI 2023. [Paper Link]

[9] Understanding and Diagnosing Deep Reinforcement Learning. International Conference on Machine Learning, ICML 2024. [Paper Link]

Filed under 21st Century Decision Making, Digital Technology and Policy Laboratory, Education, Machine Learning and Research, Master's of Public Adminstration, Policy Impact Unit, Public Policy Processes and Knowledge Systems, Research

Tags: Adversarial Reinforcement Learning, Adversarial Training, AI alignment, AI risks, AI safety, Artificial Intelligence, Deep Reinforcement Learning, Machine Learning, Machine Learning Safety, ML Safety, Reinforcement learning, Reinforcement Learning Generalization, Reinforcement Learning Safety, Responsible AI, Robust AI, Robust deep reinforcement learning, Robust Reinforcement Learning, Safe Reinforcement Learning, Trustworthy AI, Trustworthy Reinforcement Learning

No Comments »

Tracking the spread of science with machine learning

By Basil Mahfouz, on 18 November 2021

On 3-5 November 2021, I joined research professionals from across the Network for Advancing and Evaluating the Societal Impact of Science (AESIS) to discuss state of the art methods for evaluating the impact of research. Participants showcased institutional best practices, stakeholder engagement strategies, as well as how to leverage emerging data sources. In this blog, I reflect on the conversations initiated at the conference, drawing upon insights gained throughout my research at STEaPP.

Photo by Markus Spiske on Unsplash

(more…)

Filed under Digital Technology and Policy Laboratory, Education, Research

Tags: Digital Technologies and Policy, Machine Learning, research, Science, science policy

No Comments »

Machine Learning for Unlocking the Policy Impact of Transdisciplinary Research

Adversarial Attacks, Robustness and Generalization in Deep Reinforcement Learning

Tracking the spread of science with machine learning

About

Categories

Bookmarks