Inverse Reward Design

Inverse Reward Design (IRD) aims to automatically learn reward functions for reinforcement learning agents, overcoming the difficulty of manually specifying optimal reward structures for complex tasks. Current research focuses on leveraging large pre-trained models and incorporating active learning techniques, such as querying human feedback on suboptimal behaviors or using demonstrations, to efficiently infer reward functions. These methods are being applied to diverse domains, including autonomous driving and robot navigation, improving agent safety and performance while addressing challenges like reward ambiguity and generalization to unseen scenarios. The resulting advancements in IRD have significant implications for developing more robust, reliable, and human-aligned AI systems.

Papers

March 27, 2024

LORD: Large Models based Opposite Reward Design for Autonomous Driving
Xin Ye, Feng Tao, Abhirup Mallik, Burhaneddin Yaman, Liu Ren
Reinforcement Learning Autonomous Driving Reward Function Large Model Inverse Reward Design Zero Shot Reward

December 6, 2023

MIRACLE: Inverse Reinforcement and Curriculum Learning Model for Human-inspired Mobile Robot Navigation
Nihal Gunukula, Kshitij Tiwari, Aniket Bera
Inverse Reinforcement Learning Curriculum Learning Social Navigation Human Navigation Test Stimulus Inverse Reward Design Human Aware Robot Navigation

November 20, 2023

Risk-averse Batch Active Inverse Reward Design
Panagiotis Liampas
Reward Function Inverse Reward Design

September 16, 2023

Interactively Teaching an Inverse Reinforcement Learner with Limited Feedback
Rustam Zayanov, Francisco S. Melo, Manuel Lopes
Active Learning Sequential Decision Making Task Partial Feedback Causal Entropy Inverse Reward Design

June 1, 2023

Identifiability and Generalizability in Constrained Inverse Reinforcement Learning
Andreas Schlaginhaufen, Maryam Kamgarpour
Reinforcement Learning Reward Function Inverse Reinforcement Learning Stronger Generalizability Constrained Markov Decision Process Novel Identifiability Test Inverse Reward Design Inverse Constrained Reinforcement Learning

January 17, 2023

Show me what you want: Inverse reinforcement learning to automatically design robot swarms by demonstration
Ilyes Gharbi, Jonas Kuckling, David Garzón Ramos, Mauro Birattari
Inverse Reinforcement Learning Robot Swarm Textual Demonstration Collective Behavior Physical Robot Mission Specification Inverse Reward Design

October 14, 2022

Distributional Reward Estimation for Effective Multi-Agent Deep Reinforcement Learning
Jifeng Hu, Yanchao Sun, Hechang Chen, Sili Huang, haiyin piao, Yi Chang, Lichao Sun
Multi Agent Reinforcement Learning Preference Reward Inverse Reward Design

September 22, 2022

Identifiability and generalizability from multiple experts in Inverse Reinforcement Learning
Paul Rolland, Luca Viano, Norman Schuerhoff, Boris Nikolov, Volkan Cevher
Reinforcement Learning Reward Function Inverse Reinforcement Learning Stronger Generalizability Novel Identifiability Test Multiple Expert Sub Optimal Inverse Reward Design

April 11, 2022

Linguistic communication as (inverse) reward design
Theodore R. Sumers, Robert D. Hawkins, Mark K. Ho, Thomas L. Griffiths, Dylan Hadfield-Menell
Natural Language Language Supervision Inverse Task Reward Design Natural Language Communication Latent Speech Inverse Reward Design

December 30, 2021

Self Reward Design with Fine-grained Interpretability
Erico Tjoa, Guan Cuntai
Deep Reinforcement Learning Reward Function Deep RL Interpretable by Design Inverse Reward Design

Inverse Reward Design

Papers

LORD: Large Models based Opposite Reward Design for Autonomous Driving

MIRACLE: Inverse Reinforcement and Curriculum Learning Model for Human-inspired Mobile Robot Navigation

Risk-averse Batch Active Inverse Reward Design

Interactively Teaching an Inverse Reinforcement Learner with Limited Feedback

Identifiability and Generalizability in Constrained Inverse Reinforcement Learning

Show me what you want: Inverse reinforcement learning to automatically design robot swarms by demonstration

Distributional Reward Estimation for Effective Multi-Agent Deep Reinforcement Learning

Identifiability and generalizability from multiple experts in Inverse Reinforcement Learning

Linguistic communication as (inverse) reward design

Self Reward Design with Fine-grained Interpretability