Research map / Reinforcement learning / Imitation learning
In-Trajectory Inverse Reinforcement Learning: Learn Incrementally Before an Ongoing Trajectory Terminates
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MERIT-IRL, an online bi-level IRL method that learns a reward and policy from an ongoing expert trajectory using a meta-regularization prior and a single-loop OGD update that completes the trajectory with the learned policy.
Problem
Inverse reinforcement learning currently needs at least one complete expert trajectory; the paper targets learning reward and policy incrementally from an ongoing trajectory as state-action pairs arrive.
Approach
Formulate in-trajectory IRL as an online non-convex bi-level problem with a discounted meta-regularization term; the reward update compares a combined trajectory (expert prefix plus learner-rolled-out suffix) against a learner-generated trajectory, and uses one-step soft policy iteration and one-step gradient descent per time step.
Key results
- On MuJoCo, MERIT-IRL matches expert performance after observing about 40% of the trajectory (t=400), while IT-IRL needs over 90% (t=900).
- In the stock market experiment, MERIT-IRL reaches expert-level cumulative reward at t=140 (under 60% of the trajectory) with 386.70 vs expert 403.15.
- In the shooter navigation task, MERIT-IRL reaches 100% success rate at t=80 (57% of the 140-step trajectory).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add meta-regularization (ablation) | IT-IRL · HalfCheetah | cumulative reward: -386.78 -> -214.63 | Helped |
| add meta-regularization (ablation) | IT-IRL · Walker | cumulative reward: -891.79 -> -654.77 | Helped |
| add meta-regularization (ablation) | IT-IRL · Hopper | cumulative reward: -669.88 -> -476.72 | Helped |
| add meta-regularization (ablation) | IT-IRL · Stock market | cumulative reward: 256.81 -> 386.70 | Helped |
| replace the naive prefix-based reward update with the trajectory-completion reward update (ablation) | Naive MERIT-IRL · HalfCheetah | cumulative reward: -548.22 -> -214.63 | Helped |
| replace the naive prefix-based reward update with the trajectory-completion reward update (ablation) | Naive MERIT-IRL · Walker | cumulative reward: -962.42 -> -654.77 | Helped |
| replace the naive prefix-based reward update with the trajectory-completion reward update (ablation) | Naive MERIT-IRL · Hopper | cumulative reward: -691.03 -> -476.72 | Helped |
| replace the naive prefix-based reward update with the trajectory-completion reward update (ablation) | Naive MERIT-IRL · Stock market | cumulative reward: 192.49 -> 386.70 | Helped |
| use MERIT-IRL for in-trajectory learning | IT-IRL and naive baselines · shooter ongoing trajectory | success rate: 100% at t=80 | Helped |
| add meta-regularization (ablation) | IT-IRL · MuJoCo target velocity tasks | cumulative reward | Helped |
| use the trajectory-completion reward update instead of the naive prefix update (ablation) | Naive MERIT-IRL and Naive IT-IRL · MuJoCo target velocity tasks | cumulative reward | Helped |
| add meta-regularization (ablation) | IT-IRL and other in-trajectory baselines · Stock market | cumulative reward | Helped |
Surprising findings
- Naive in-trajectory variants that compare trajectory prefixes fail to imitate the expert even after the full trajectory is observed.
Limitations
- Theoretical guarantees rely on assumptions such as bounded reward gradients and uniform ergodicity of the Markov chain.
- Regret bounds are local for nonlinear rewards; global sub-linear regret is only shown for linear rewards.
Experimental setup
Datasets: MuJoCo HalfCheetah, Walker, Hopper target-velocity tasks Dow Jones 30 constituent stocks 2021 via FinRL Gazebo shooter navigation simulator
Benchmarks: MuJoCo target velocity (HalfCheetah, Walker, Hopper) FinRL stock market Gazebo active shooter navigation
Metrics: cumulative ground-truth reward success rate local regret (theoretical)
Baselines: IT-IRL Naive MERIT-IRL Naive IT-IRL Hindsight (meta-regularized ML-IRL)
Builds on: maximum likelihood IRL model-agnostic meta-learning soft Q-learning soft actor-critic online non-convex optimization local regret
Closest papers by meaning
- Recent Advancements in Inverse Reinforcement Learning (AAAI 2024 · similarity 0.89)
- Is Inverse Reinforcement Learning Harder than Standard Reinforcement Learning? A Theoretical Perspective (ICML 2024 · similarity 0.89)
- Offline Inverse RL: New Solution Concepts and Provably Efficient Algorithms (ICML 2024 · similarity 0.89)
- A Geometric Perspective on Reward Function Updates in Inverse Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- How does Inverse RL Scale to Large State Spaces? A Provably Efficient Approach (NeurIPS 2024 · similarity 0.88)
- A Unified Linear Programming Framework for Offline Reward Learning from Human Demonstrations and Feedback (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.