Natural Language Specification of Reinforcement Learning Policies Through Differentiable Decision Trees

Tambwekar, Pradyumna; Silva, Andrew; Gopalan, Nakul; Gombolay, Matthew

doi:10.1109/lra.2023.3268593

Cited by 4 publications

(1 citation statement)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Prior work has proposed differentiable decision trees (DDTs) (Suárez and Lutsko, 1999;Silva and Gombolay, 2021;Paleja et al, 2020;Tambwekar et al, 2023) -a neural network architecture that takes the topology of a decision tree (DT). Similar to a decision tree, DDTs contain decision nodes and leaf nodes; however, each decision node within the DDT utilizes a sigmoid activation function (i.e., a "soft" decision) instead of a Boolean decision (i.e., a "hard" decision).…”

Section: Differentiable Decision Trees (Ddts)mentioning

confidence: 99%

Learning Interpretable, High-Performing Policies for Autonomous Driving

Paleja¹,

Niu²,

Silva³

et al. 2022

Robotics: Science and Systems XVIII

View full text Add to dashboard Cite

Gradient-based approaches in reinforcement learning (RL) have achieved tremendous success in learning policies for autonomous vehicles. While the performance of these approaches warrants real-world adoption, these policies lack interpretability, limiting deployability in the safety-critical and legally-regulated domain of autonomous driving (AD). AD requires interpretable and verifiable control policies that maintain high performance. We propose Interpretable Continuous Control Trees (ICCTs), a tree-based model that can be optimized via modern, gradient-based, RL approaches to produce highperforming, interpretable policies. The key to our approach is a procedure for allowing direct optimization in a sparse decisiontree-like representation. We validate ICCTs against baselines across six domains, showing that ICCTs are capable of learning interpretable policy representations that parity or outperform baselines by up to 33% in AD scenarios while achieving a 300x-600x reduction in the number of policy parameters against deep learning baselines. Furthermore, we demonstrate the interpretability and utility of our ICCTs through a 14-car physical robot demonstration. 1) We propose Interpretable Continuous Control Trees (IC-CTs), a novel tree-based model that can be optimized via gradient descent with modern RL algorithms to produce high-performance, interpretable continuous control policies. We provide several extensions to prior DDT i x k −b i > 0). As each leaf node is represented as a probability mass function over output classes in prior work, leaf nodes, l, must be modified to produce

show abstract

Section: Differentiable Decision Trees (Ddts)mentioning

confidence: 99%

Learning Interpretable, High-Performing Policies for Autonomous Driving

Paleja¹,

Niu²,

Silva³

et al. 2022

Robotics: Science and Systems XVIII

View full text Add to dashboard Cite

show abstract

Effects of explainable artificial intelligence in neurology decision support

Gombolay,

Silva,

Schrum

et al. 2024

Ann Clin Transl Neurol

Self Cite

View full text Add to dashboard Cite

ObjectiveArtificial intelligence (AI)‐based decision support systems (DSS) are utilized in medicine but underlying decision‐making processes are usually unknown. Explainable AI (xAI) techniques provide insight into DSS, but little is known on how to design xAI for clinicians. Here we investigate the impact of various xAI techniques on a clinician's interaction with an AI‐based DSS in decision‐making tasks as compared to a general population.MethodsWe conducted a randomized, blinded study in which members of the Child Neurology Society and American Academy of Neurology were compared to a general population. Participants received recommendations from a DSS via a random assignment of an xAI intervention (decision tree, crowd sourced agreement, case‐based reasoning, probability scores, counterfactual reasoning, feature importance, templated language, and no explanations). Primary outcomes included test performance and perceived explainability, trust, and social competence of the DSS. Secondary outcomes included compliance, understandability, and agreement per question.ResultsWe had 81 neurology participants with 284 in the general population. Decision trees were perceived as the more explainable by the medical versus general population (P < 0.01) and as more explainable than probability scores within the medical population (P < 0.001). Increasing neurology experience and perceived explainability degraded performance (P = 0.0214). Performance was not predicted by xAI method but by perceived explainability.InterpretationxAI methods have different impacts on a medical versus general population; thus, xAI is not uniformly beneficial, and there is no one‐size‐fits‐all approach. Further user‐centered xAI research targeting clinicians and to develop personalized DSS for clinicians is needed.

show abstract