Entangled bisimulation improves policy learning from visual input.
problem Learning generalizeable policies from visual input in the presence of visual distractions.
method Proposes entangled bisimulation, a bisimulation metric for continuous state and action spaces.
result Entangled bisimulation improves policy learning on the Distracting Control Suite (DCS).
Action-bisimulation learns long-horizon controllability for reinforcement learning.
problem Learning relevant state features in high-dimensional observations for robust reinforcement learning.
method Action-bisimulation encoding, inspired by bisimulation invariance, extends single-step controllability to multi-step.
result Action-bisimulation pretraining improves sample efficiency in various environments.
Estimates bisimulation metrics from sample streams, not full transition models.
problem Estimating Markov chain metrics from limited sample data.
method Stochastic optimization using linear programming and primal-dual method.
result Validated through empirical evaluations, providing sample complexity guarantees.
Method learns representations invariant to task-irrelevant details in reinforcement learning tasks.
problem Learning representations that are invariant to task-irrelevant details in reinforcement learning.
method Uses bisimulation metrics to learn robust latent representations that encode only task-relevant information.
result Demonstrates SOTA performance in modified visual MuJoCo tasks and a first-person driving task.
We present new algorithms for computing and approximating bisimulation metrics in Markov Decision Processes (MDPs). Bisimulation metrics are an elegant formalism that capture behavioral equivalence between states and provide strong theoretical guarantees on differences in optimal behaviour. Unfortunately, their computa…
Enhances RL in partially observable, noisy environments by uncovering causal states.
problem Making decisions based on incomplete and noisy observations in partially observable Markov decision processes (P2OMDPs). method Causal State Representation under Asynchronous Diffusion Model (CaDiff) framework, incorporating a novel asynchronous diffusion model (ADM) and a new bisimulation metric.
result Enhances returns by at least 14.18% compared to baselines on Roboschool tasks.
The paper formalizes robot environments using topological concepts.
problem Determining indistinguishable environments for robots.
method Formalizing robot systems as topological dynamical systems, using covering maps and bisimulation.
result Covering maps can witness indistinguishability of environments under general conditions.
Deep neural network learns discrete state abstractions for efficient planning.
problem Efficient sequential decision making in large state spaces.
method Information bottleneck method for learning approximate bisimulations using deep neural encoders and action-conditioned HMM.
result Trained method efficiently plans for unseen goals in multi-goal reinforcement learning.
New framework for efficient optimal transport distances between Markov chains.
problem Efficient computation of optimal transport distances between Markov chains.
method Developed a new perspective on optimal transport distances using discounted occupancy couplings and linear programming.
result Introduced Sinkhorn Value Iteration (SVI) for efficient calculation of optimal transport distances.
CEA augments reinforcement learning by generating counterfactual experiences.
problem Challenges in reinforcement learning, especially out-of-distribution and inefficient exploration.
method CEA uses variational autoencoders to model state transitions and introduces randomness for non-stationarity. It expands learning data through counterfactual inference.
result CEA outperforms SOTA algorithms in diverse environments.
Many dynamical systems exhibit similar structure, as often captured by hand-designed simplified models that can be used for analysis and control. We develop a method for learning to correspond pairs of dynamical systems via a learned latent dynamical system. Given trajectory data from two dynamical systems, we learn a …
Intelligent agents can cope with sensory-rich environments by learning task-agnostic state abstractions. In this paper, we propose an algorithm to approximate causal states, which are the coarsest partition of the joint history of actions and observations in partially-observable Markov decision processes (POMDP). Our m…
Many reinforcement learning (RL) tasks provide the agent with high-dimensional observations that can be simplified into low-dimensional continuous states. To formalize this process, we introduce the concept of a DeepMDP, a parameterized latent space model that is trained via the minimization of two tractable losses: pr…
In this work we present a novel approach for transfer-guided exploration in reinforcement learning that is inspired by the human tendency to leverage experiences from similar encounters in the past while navigating a new task. Given an optimal policy in a related task-environment, we show that its bisimulation distance…
The paper proposes a principle for dynamically adjusting the granularity of reinforcement learning abstractions.
problem Lack of general principles for dynamically adjusting the granularity of reinforcement learning abstractions.
method The paper proposes a principle based on rate-distortion theory, formalized through a performance certificate decomposing value error into learning and abstraction error bounds.
result Soft state-action abstractions can achieve near-optimal performance under substantial lossy compression of state and action information.