This study compares global vs local observation and action representations for DRL in RTS games.
problem Improving Deep Reinforcement Learning performance in RTS games.
method Comparing two observation and action representations in μRTS.
result Local representation outperforms global representation in resource harvesting tasks.
The paper formalizes criteria for non-spurious and disentangled representations using causal methods.
problem Formalizing criteria for non-spurious and disentangled representations in representation learning.
method Causal perspective, counterfactual quantities, observable consequences of causal assertions.
result Computable metrics for assessing representation learning based on observed data.
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…
Representation learning is a central challenge across a range of machine learning areas. In reinforcement learning, effective and functional representations have the potential to tremendously accelerate learning progress and solve more challenging problems. Most prior work on representation learning has focused on gene…
We would like to learn a representation of the data which decomposes an observation into factors of variation which we can independently control. Specifically, we want to use minimal supervision to learn a latent representation that reflects the semantics behind a specific grouping of the data, where within a group the…
New method learns stochastic process representations without exact reconstruction.
problem Learning exact representations of high-dimensional noisy stochastic processes.
method CReSP framework for contrastive learning of stochastic processes.
result Effective for learning representations of various stochastic processes.
Unsupervised representation learning has succeeded with excellent results in many applications. It is an especially powerful tool to learn a good representation of environments with partial or noisy observations. In partially observable domains it is important for the representation to encode a belief state, a sufficie…
Disentangled representations, where the higher level data generative factors are reflected in disjoint latent dimensions, offer several benefits such as ease of deriving invariant representations, transferability to other tasks, interpretability, etc. We consider the problem of unsupervised learning of disentangled rep…
Temporal-difference (TD) networks are a class of predictive state representations that use well-established TD methods to learn models of partially observable dynamical systems. Previous research with TD networks has dealt only with dynamical systems with finite sets of observations and actions. We present an algorithm…
New method identifies causal variables from partially observed data.
problem Learning from unpaired observations with instance-dependent partial observability.
method Proposes two methods enforcing sparsity in the inferred representation.
result Establishes two identifiability results for linear and piecewise linear mixing functions.
FSRM method improves treatment effect estimation from observational data.
problem Estimating treatment effects from observational data with missing counterfactual outcomes and selection bias.
method FSRM method based on deep representation learning and matching, which maps covariate space into a selective, nonlinear, and balanced representation space.
result FSRM method outperforms state-of-the-art methods in estimating treatment effects.
Method models other agents' behaviors without requiring direct observation.
problem Understanding and interacting effectively with other agents in reinforcement learning.
method Extracts representations from local observations of the controlled agent using encoder-decoder architectures.
result The method achieves higher returns than baseline methods in multi-agent environments.
Animals need to devise strategies to maximize returns while interacting with their environment based on incoming noisy sensory observations. Task-relevant states, such as the agent's location within an environment or the presence of a predator, are often not directly observable but must be inferred using available sens…
A method for self-supervised representation learning in partially observable environments.
problem Sparse rewards and stochasticity in partially observable environments.
method World model in latent space for estimating missing information.
result Significant improvement in exploration compared to prior work.
Framework for causal discovery using multi-modal data.
problem Failure of representation learning in causal tasks.
method Statistical and computational framework combining representation learning and causal inference.
result Effective use of observational and perturbational data for causal discovery.
New method learns useful disentangled representations from weakly labeled data.
problem Learning useful representations from weakly labeled data.
method Model pairs of non-i.i.d. images, learn disentangled representations without requiring annotation.
result Learn disentangled representations reliably from pairs of images without requiring group, individual factor, or number of changed factors annotation.
Study learns state representations from observations for control, proving guarantees.
problem Learning state representations from high-dimensional observations for control.
method Cost-driven approach, learning latent state model to predict costs.
result Proves finite-sample guarantees for near-optimal state representation and controller.
Method estimates causal effects from incremental data, overcoming missing data challenges.
problem Estimating causal effects from non-stationary, incrementally available observational data.
method Continual Causal Effect Representation Learning
result Method achieves continual causal effect estimation without compromising original data.
iTimER learns from reconstruction errors to represent irregularly sampled time series.
problem Learning from irregularly sampled time series with missing data.
method iTimER models reconstruction errors as a proxy for unobserved values, using a mixup strategy and a Wasserstein metric.
result iTimER outperforms state-of-the-art methods in classification, interpolation, and forecasting tasks.
Predictive State Representations (PSRs) are an expressive class of models for controlled stochastic processes. PSRs represent state as a set of predictions of future observable events. Because PSRs are defined entirely in terms of observable data, statistically consistent estimates of PSR parameters can be learned effi…
Proposes FunNoL for better curve classification and reconstruction in multivariate functional data.
problem Linear methods fail to capture nonlinear structures in multivariate functional data.
method Functional nonlinear learning (FunNoL) method using nonlinear mapping.
result FunNoL outperforms FPCA in curve classification and reconstruction, especially in multivariate settings.
Efficiently extracts linear dynamics from complex observations.
problem Learning policies directly from rich, high-dimensional observations.
method Modeling linear dynamics in a hidden subspace and developing an efficient algorithm.
result Successfully extracts linear dynamics from rich observations.
This paper evaluates various representations for robotics tasks, improving performance in lifting, stacking, and pushing.
problem Improving data-efficiency in reinforcement learning for robotics with limited data.
method Systematic evaluation of common representations in three robotics tasks: lifting, stacking, and pushing.
result Some representations can perform as well as simulator states as agent inputs, challenging common intuitions.
This paper provides an analysis of the tradeoff between asymptotic bias (suboptimality with unlimited data) and overfitting (additional suboptimality due to limited data) in the context of reinforcement learning with partial observability. Our theoretical analysis formally characterizes that while potentially increasin…
Recently there has been a significant interest in learning disentangled representations, as they promise increased interpretability, generalization to unseen scenarios and faster learning on downstream tasks. In this paper, we investigate the usefulness of different notions of disentanglement for improving the fairness…
Many real-world sequential decision-making problems can be formulated as optimal control with high-dimensional observations and unknown dynamics. A promising approach is to embed the high-dimensional observations into a lower-dimensional latent representation space, estimate the latent dynamics model, then utilize this…
Kernel embeddings help estimate causal effects from observational data.
problem Estimating causal effects from observational data with confounding variables.
method Kernel embeddings in reproducing kernel Hilbert spaces (RKHS).
result Robust nonparametric framework for causal inference.
ETC learns minimal representations for reinforcement learning in POMDPs.
problem Sample complexity challenges in reinforcement learning for POMDPs.
method ETC learns low-dimensional features and embeddings at two levels, optimizing policy.
result ETC achieves polynomial sample complexity for POMDPs with low-rank transition kernels.
Develops a smooth operator framework for analyzing neural network representations.
problem Analyzing the geometry of feedforward neural network representations.
method Introduces a smooth operator-theoretic approach based on diffusion Markov operators derived from feature clouds.
result Establishes a stable operator-geometric framework for tracking training, width, and perturbation stability.
Unpaired multi-domain causal representation learning is possible with sufficient conditions.
problem Learning shared causal representation from unpaired data across domains.
method Identify sufficient conditions for joint distribution and shared causal graph recovery.
result Practical method to recover shared latent causal graph from marginal distributions.
New algorithm improves reinforcement learning from partial observations.
problem Inferior performance of algorithms in real-world reinforcement learning due to partial observability.
method Representation-based approach to POMDPs, leading to a tractable algorithm.
result Empirically demonstrates superior performance with partial observations.
We present the perceptor gradients algorithm -- a novel approach to learning symbolic representations based on the idea of decomposing an agent's policy into i) a perceptor network extracting symbols from raw observation data and ii) a task encoding program which maps the input symbols to output actions. We show that t…
New method uses bi-level optimization to learn useful representations for imitation learning.
problem Learning useful representations for multiple tasks in imitation learning settings.
method Formulates representation learning as a bi-level optimization problem.
result Bi-level optimization framework provides sample complexity benefits for imitation learning.
iCITRIS learns causal variables from interactive systems with instantaneous effects.
problem Identifying causal variables from temporal sequences with instantaneous effects.
method iCITRIS method for causal representation learning that handles instantaneous effects in intervened temporal sequences.
result iCITRIS accurately identifies causal variables and their causal graph from three interactive system datasets.
Defines complexity measure for neural networks and feature representations, revealing scaling patterns.
problem Understanding the nonlinearity and dimensionality of neural network computations and feature representations.
method Introduces complexity and effective dimension measures, investigates their dynamics during training, and analyzes their scaling properties.
result Power law scaling of complexity and effective dimension during training, revealing hidden structure of datasets.
Combines observational and randomized data to estimate treatment effects.
problem Estimating heterogeneous treatment effects using only observational data is biased.
method Two-step framework: learn shared structure from observational data, then data-specific structures from randomized data.
result Combining observational and randomized data improves treatment effect estimation.
Recurrent neural networks (RNNs) are an effective representation of control policies for a wide range of reinforcement and imitation learning problems. RNN policies, however, are particularly difficult to explain, understand, and analyze due to their use of continuous-valued memory vectors and observation features. In …
Finding a generally accepted formal definition of a disentangled representation in the context of an agent behaving in an environment is an important challenge towards the construction of data-efficient autonomous agents. Higgins et al. recently proposed Symmetry-Based Disentangled Representation Learning, a definition…
VED framework learns low-dimensional latent representations of physical systems.
problem Learning latent representations of complex physical systems.
method Variational Encoder-Decoder (VED) framework with KL divergence and covariance regularization.
result VED achieves lower-dimensional latent representations with improved feature disentanglement.
The paper proposes a method to precisely decompose confounders and estimate treatment effects.
problem Estimating treatment effects from observational data with confounder identification and balancing.
method Learning decomposed representations to identify and balance confounders and non-confounders.
result The method achieves more precise treatment effect estimation than existing methods.
New method identifies key genes affecting phenotypes in biological systems.
problem Identifying genes that drive specific phenotypes in complex biological systems.
method Data-driven observability decomposition using Koopman operators.
result Koopman operator representation identifies genes that drive phenotypes.
New method identifies causal relationships without strong assumptions.
problem Causal Representation Learning (CRL) is ill-posed due to representation and causal discovery issues.
method Identifiability based on grouping of observational variables, self-supervised estimation framework.
result Practical identifiability conditions without temporal structure, interventions, or weak supervision.
In recent years, deep generative models have been shown to 'imagine' convincing high-dimensional observations such as images, audio, and even video, learning directly from raw data. In this work, we ask how to imagine goal-directed visual plans -- a plausible sequence of observations that transition a dynamical system …
Paper learns data-driven organ matching rules from observational data.
problem Tackles organ transplantation compatibility using observational data.
method Representation learning to cluster donors and apply recipient transformations.
result Model outperforms human experts in predicting transplant outcomes.
Motivated by the study of the interrelation between functorial and algebraic quantum field theory, we point out that on any locally trivial bundle of compact groups, representations up to homotopy are enough to separate points by means of the associated representations in cohomol- ogy. Furthermore, we observe that the …
New method identifies latent causal variables from observed data, overcoming indeterminacies.
problem Identifying latent causal variables from observed data, especially when latent variables are weight-variant.
method Introduces a novel identifiability condition for latent causal models, proposing SuaVE method.
result Identifies latent causal variables up to trivial permutation and scaling, demonstrating consistency and efficacy.
Earth observation embeddings can convert discrete biome maps into continuous representations that better capture ecological variation.
problem Biome maps impose categorical boundaries that compress continuous variation in biotic communities.
method Fit a linear classifier on Earth observation embeddings to predict biome labels.
result Continuous biome representation outperforms discrete biome labels for predicting species occurrence.
The paper provides bounds and methods for estimating causal effects from observational data.
problem Estimating individual-level causal effects from non-experimental data.
method Generalization bounds, sample re-weighting, representation learning algorithms.
result The proposed methods reduce treatment group distances and improve estimation accuracy.