[Paper Review] Meta-Reinforcement Learning in Broad and Non-Parametric Environments

Zhenshan Bing, Lukas Knak+3 more|arXiv (Cornell University)|Aug 8, 2021

Domain Adaptation and Few-Shot Learning20 references4 citations

TL;DR

This paper proposes TIGR, a meta-reinforcement learning algorithm that enables zero-shot adaptation to non-parametric and non-stationary tasks using a Gaussian variational autoencoder (VAE) for task inference and gated recurrent units (GRUs) to model task dynamics. It achieves 3–10× faster sample efficiency and superior asymptotic performance compared to state-of-the-art methods in non-parametric environments with dynamic task shifts.

ABSTRACT

Recent state-of-the-art artificial agents lack the ability to adapt rapidly to new tasks, as they are trained exclusively for specific objectives and require massive amounts of interaction to learn new skills. Meta-reinforcement learning (meta-RL) addresses this challenge by leveraging knowledge learned from training tasks to perform well in previously unseen tasks. However, current meta-RL approaches limit themselves to narrow parametric task distributions, ignoring qualitative differences between tasks that occur in the real world. In this paper, we introduce TIGR, a Task-Inference-based meta-RL algorithm using Gaussian mixture models (GMM) and gated Recurrent units, designed for tasks in non-parametric environments. We employ a generative model involving a GMM to capture the multi-modality of the tasks. We decouple the policy training from the task-inference learning and efficiently train the inference mechanism on the basis of an unsupervised reconstruction objective. We provide a benchmark with qualitatively distinct tasks based on the half-cheetah environment and demonstrate the superior performance of TIGR compared to state-of-the-art meta-RL approaches in terms of sample efficiency (3-10 times faster), asymptotic performance, and applicability in non-parametric environments with zero-shot adaptation.

Motivation & Objective

• To address the limitation of existing meta-RL methods that assume narrow, parametric, and stationary task distributions.
• To enable rapid adaptation to qualitatively different, non-parametric tasks—such as running forward, backward, or jumping—without requiring task-specific fine-tuning.
• To support zero-shot adaptation to non-stationary task changes that occur during inference, not just at test time.
• To combine model-free, off-policy reinforcement learning with unsupervised task inference for improved data efficiency and generalization.
• To establish a benchmark for non-parametric task generalization in robotic control environments like half-cheetah and ant.

Proposed method

• TIGR employs a Gaussian variational autoencoder (VAE) to explicitly model multi-modal task distributions, capturing qualitative differences between tasks.
• It decouples task inference from policy training by pre-training the VAE on reconstructed MDPs using an unsupervised reconstruction objective.
• A gated recurrent unit (GRU) processes recent transition history to infer task identity in real time, enabling dynamic adaptation to non-stationary task changes.
• The policy is trained off-policy using SAC (Soft Actor-Critic) with task-conditional latent codes inferred via the VAE and GRU.
• Task inference is performed at each timestep using features from the most recent transition history, allowing immediate adaptation to new task modes.
• The method uses a multi-component VAE with 8 clusters to represent distinct task types, such as different movement objectives in the half-cheetah environment.

Experimental results

Research questions

• RQ1 Can a meta-RL agent generalize to qualitatively different, non-parametric tasks such as running forward vs. backward or jumping vs. turning?
• RQ2 Can the agent achieve zero-shot adaptation to non-stationary task changes that occur during inference, without requiring retraining?
• RQ3 Does the use of a VAE-based generative model for task inference improve sample efficiency and asymptotic performance compared to Gaussian task embedding methods?
• RQ4 How does TIGR perform in environments with high non-parametric variability compared to state-of-the-art meta-RL baselines?
• RQ5 Can unsupervised reconstruction of MDPs during meta-training effectively support downstream policy learning in complex, diverse task distributions?

Key findings

• TIGR achieves 3–10 times faster sample efficiency than state-of-the-art methods like PEARL and MAML in non-parametric environments.
• The algorithm demonstrates substantially improved asymptotic performance on non-parametric tasks, particularly in zero-shot adaptation scenarios.
• TIGR successfully generalizes to eight distinct non-parametric tasks in the half-cheetah environment, including forward/backward running, goal-reaching, and flipping.
• The use of a multi-modal VAE enables accurate task inference across qualitatively different behaviors, even when task changes occur dynamically during inference.
• The zero-shot adaptation mechanism, based on real-time GRU-based task inference from recent transitions, allows immediate response to task shifts without retraining.
• Empirical results show that TIGR outperforms existing methods in both sample efficiency and final performance across multiple non-parametric benchmarks, including modified half-cheetah and ant environments.