Periodic Reporting for period 1 - ixAutoML (Interactive and Explainable Human-Centered AutoML)
Reporting period: 2022-12-01 to 2025-05-31
The research conducted under this ERC Starting Grant project aims to advance the field of Automated Machine Learning (AutoML) by focusing on three primary objectives: (i) explaining static AutoML, (ii) explaining dynamic AutoML policies, including those leveraging Reinforcement Learning (RL), and (iii) creating a human-centered, interactive AutoML paradigm. By addressing these objectives, we envision an ecosystem where AutoML systems are trusted, effective, and accessible, empowering users to bring their expertise into the optimization process and improve its overall efficiency and reliability. The most transformative aspect of this project is its focus on creating a human-centered, interactive AutoML paradigm. Traditional AutoML approaches often exclude users from the decision-making and optimization loop. In contrast, we envision a paradigm where users actively contribute their knowledge and preferences. The overarching vision of this project is to reshape the role of AutoML in machine learning practice, emphasizing trust, efficiency, and inclusivity of human developers. By increasing the explainability of both static and dynamic AutoML systems, we aim to build confidence among users, especially in high-stakes applications where transparency is paramount. Trust in these systems is further reinforced by our efforts to create explainability and user-guided frameworks, ensuring that the decision-making process aligns with human intuition and expertise.
The ixAutoML project has introduced several novel methodologies to address the challenges of trust and accessibility in AutoML.
A key methodological innovation is the development of symbolic explanation techniques for hyperparameter optimization based on performance data collected via Bayesian Optimization, making complex AutoML decisions comprehensible for users. In addition, we developed a new variant of Bayesian Optimization (BO) that re-interprets the Bayesian Execution Algorithm (BAX) to obtain high-quality partial dependence plots (PDPs). By interleaving BAX with traditional BO, we can show that the optimization performance does not decrease while obtaining much better PDPs.
We are the first to propose a systematic approach for studying the hyperparameter landscapes of reinforcement learning algorithms dynamically during policy training. Since at each decision point, we could spawn a large number of new policy trainings, we are facing a combinatorial problem. We propose to store checkpoints and continue policy learning from there until convergence. By choosing the best-performing policies out of all hyperparameter configurations, we can go back to a checkpoint and spawn new hyperparameter configuration runs from there. This technique is overall very efficient and allows a deeper understanding of the landscapes.
To improve the efficiency, robustness, and soundness of AutoRL systems, we propose a recipe for efficient RL research, incl. training test splits on different levels, designing configuration spaces, deciding on the correct optimization approach, settling on an important cost metric, running HPO for several seeds and evaluate on new seeds. Although, maybe surprisingly trivial, we were able to show that results in comparing RL and AutoRL algorithms improved in quality substantially.
To enable interaction between users and AutoML packages, we propose to bias hyperparameter optimization towards users' expectations (e.g. preferred learning rates). In contrast to previous approaches, we proposed a simple-to-implement generic approach based on re-weighting the acquisition function w.r.t. the user prior that can outperform complex previous approaches and keeps the theoretical guarantees for BO.
Furthermore, we have established an interactive framework for hyperparameter optimization (HPO), leveraging preference learning to decide between examples of Pareto fronts. The learned preferences can then be used to guide HPO towards preferred shapes of Pareto Fronts. This allows human users to actively guide the HPO process in multi-objective settings, creating a seamless integration of human intuition and algorithmic efficiency.
A key methodological innovation is the development of symbolic explanation techniques for hyperparameter optimization based on performance data collected via Bayesian Optimization, making complex AutoML decisions comprehensible for users. In addition, we developed a new variant of Bayesian Optimization (BO) that re-interprets the Bayesian Execution Algorithm (BAX) to obtain high-quality partial dependence plots (PDPs). By interleaving BAX with traditional BO, we can show that the optimization performance does not decrease while obtaining much better PDPs.
We are the first to propose a systematic approach for studying the hyperparameter landscapes of reinforcement learning algorithms dynamically during policy training. Since at each decision point, we could spawn a large number of new policy trainings, we are facing a combinatorial problem. We propose to store checkpoints and continue policy learning from there until convergence. By choosing the best-performing policies out of all hyperparameter configurations, we can go back to a checkpoint and spawn new hyperparameter configuration runs from there. This technique is overall very efficient and allows a deeper understanding of the landscapes.
To improve the efficiency, robustness, and soundness of AutoRL systems, we propose a recipe for efficient RL research, incl. training test splits on different levels, designing configuration spaces, deciding on the correct optimization approach, settling on an important cost metric, running HPO for several seeds and evaluate on new seeds. Although, maybe surprisingly trivial, we were able to show that results in comparing RL and AutoRL algorithms improved in quality substantially.
To enable interaction between users and AutoML packages, we propose to bias hyperparameter optimization towards users' expectations (e.g. preferred learning rates). In contrast to previous approaches, we proposed a simple-to-implement generic approach based on re-weighting the acquisition function w.r.t. the user prior that can outperform complex previous approaches and keeps the theoretical guarantees for BO.
Furthermore, we have established an interactive framework for hyperparameter optimization (HPO), leveraging preference learning to decide between examples of Pareto fronts. The learned preferences can then be used to guide HPO towards preferred shapes of Pareto Fronts. This allows human users to actively guide the HPO process in multi-objective settings, creating a seamless integration of human intuition and algorithmic efficiency.
Explaining Static AutoML
Static AutoML methods are widely used but often perceived as black-box systems, which can limit their adoption. Our research has made strides in enhancing the transparency and explainability of these approaches. One notable achievement is the development of frameworks such as Enhancing Explainability of Hyperparameter Optimization via Bayesian Algorithm Execution. This work guides the AutoML optimization process so that more reliable and trustworthy explanations can be derived post hoc without investigating further configurations of the ML system.
Additionally, we introduced Symbolic Explanations for Hyperparameter Optimization, a method that generates interpretable, symbolic representations of relationships between hyperparameters and their effects on model performance. This approach empowers users to comprehend the rationale behind AutoML-generated configurations, bridging the gap between complex optimization algorithms and machine learning expertise. We believe that this will make AutoML also more interesting as a tool to derive theories in ML learning processes.
Explaining Dynamic AutoML Policies
Dynamic AutoML systems, including those based on RL, adapt their policies in real time based on evolving data and changing objectives. These systems pose unique challenges regarding explainability, as their decisions are inherently more complex and context-dependent. Our work on AutoRL Hyperparameter Landscapes provides fundamental insights into the structure and behavior of RL hyperparameter spaces, enabling more informed tuning strategies.
In Hyperparameters in Reinforcement Learning and How to Tune Them, we were the first to show how difficult the tuning of RL algorithms is. We synthesized practical guidelines for optimizing RL hyperparameters, offering practitioners a robust framework for improving model performance.
A Human-Centered Interactive AutoML Paradigm
A key aspect of our vision is to empower users to bring their knowledge into the optimization loop. Approaches like πBO: Augmenting Acquisition Functions with User Beliefs for Bayesian Optimization demonstrate how user expertise on promising hyperparameter configurations can be seamlessly integrated into the optimization processes, improving performance and the perceived reliability of the system. This work is complemented by PriorBand: Practical Hyperparameter Optimization in the Age of Deep Learning, introducing a novel approach leveraging prior knowledge to enhance optimization efficiency for expensive deep neural networks.
We also developed Interactive Hyperparameter Optimization in Multi-Objective Problems via Preference Learning, a framework that learns user preferences to guide optimization processes in complex, multi-objective settings. This is particularly important in cases where accuracy is just as important as, e.g. fairness, energy consumption, or latency. Our approach enables users to guide hyperparameter optimization in these multi-objective settings without understanding complex multi-objective metrics by ranking a few different possible Pareto fronts.
Static AutoML methods are widely used but often perceived as black-box systems, which can limit their adoption. Our research has made strides in enhancing the transparency and explainability of these approaches. One notable achievement is the development of frameworks such as Enhancing Explainability of Hyperparameter Optimization via Bayesian Algorithm Execution. This work guides the AutoML optimization process so that more reliable and trustworthy explanations can be derived post hoc without investigating further configurations of the ML system.
Additionally, we introduced Symbolic Explanations for Hyperparameter Optimization, a method that generates interpretable, symbolic representations of relationships between hyperparameters and their effects on model performance. This approach empowers users to comprehend the rationale behind AutoML-generated configurations, bridging the gap between complex optimization algorithms and machine learning expertise. We believe that this will make AutoML also more interesting as a tool to derive theories in ML learning processes.
Explaining Dynamic AutoML Policies
Dynamic AutoML systems, including those based on RL, adapt their policies in real time based on evolving data and changing objectives. These systems pose unique challenges regarding explainability, as their decisions are inherently more complex and context-dependent. Our work on AutoRL Hyperparameter Landscapes provides fundamental insights into the structure and behavior of RL hyperparameter spaces, enabling more informed tuning strategies.
In Hyperparameters in Reinforcement Learning and How to Tune Them, we were the first to show how difficult the tuning of RL algorithms is. We synthesized practical guidelines for optimizing RL hyperparameters, offering practitioners a robust framework for improving model performance.
A Human-Centered Interactive AutoML Paradigm
A key aspect of our vision is to empower users to bring their knowledge into the optimization loop. Approaches like πBO: Augmenting Acquisition Functions with User Beliefs for Bayesian Optimization demonstrate how user expertise on promising hyperparameter configurations can be seamlessly integrated into the optimization processes, improving performance and the perceived reliability of the system. This work is complemented by PriorBand: Practical Hyperparameter Optimization in the Age of Deep Learning, introducing a novel approach leveraging prior knowledge to enhance optimization efficiency for expensive deep neural networks.
We also developed Interactive Hyperparameter Optimization in Multi-Objective Problems via Preference Learning, a framework that learns user preferences to guide optimization processes in complex, multi-objective settings. This is particularly important in cases where accuracy is just as important as, e.g. fairness, energy consumption, or latency. Our approach enables users to guide hyperparameter optimization in these multi-objective settings without understanding complex multi-objective metrics by ranking a few different possible Pareto fronts.