Temporal difference learning explained through gradient splitting, improving convergence times.
problem Learning value functions in Markov Decision Processes with linear approximations.
method Interpreting TD learning as gradient splitting and applying convergence proofs from gradient descent.
result Improved convergence times for TD learning, especially with a minor variation.
Quantile TD learning outperforms classical TD learning for value estimation.
problem Temporal-difference learning in reinforcement learning.
method Quantile Temporal-Difference Learning (QTD) for policy evaluation.
result QTD offers superior performance to classical TD learning, even in tabular settings.
Temporal-difference and Q-learning learn feature representations that converge to optimal ones.
problem Understanding how feature representations evolve in temporal-difference and Q-learning with neural networks.
method Mean-field theory applied to overparameterized two-layer neural networks.
result The feature representation converges to the optimal one, generalizing previous results.
Here, we provide a supplementary material for Takayuki Osogami, "Uncorrected least-squares temporal difference with lambda-return," which appears in {\it Proceedings of the 34th AAAI Conference on Artificial Intelligence} (AAAI-20).
New methods improve temporal difference learning for policy evaluation in Markov decision processes.
problem Improving temporal difference learning for policy evaluation in Markov decision processes.
method Introduced variance-reduced forms of stochastic approximation to achieve non-asymptotic, instance-dependent optimality.
result Temporal difference learning is strictly suboptimal, but variance-reduced forms achieve optimality up to logarithmic factors.
Quantile Temporal-Difference learning proved convergent with proof.
problem Lack of theoretical understanding of QTD despite empirical success.
method Proof of convergence using stochastic approximation and non-smooth analysis.
result QTD converges to fixed points with probability 1.
TD-Flow improves long-term predictions in agent learning.
problem Cumulative errors in step-by-step inference of future states.
method Leverages flow-matching techniques and a novel Bellman equation to learn accurate geometric horizon models.
result Significantly reduces errors at long horizons compared to prior methods.
Temporal difference learning and Residual Gradient methods are the most widely used temporal difference based learning algorithms; however, it has been shown that none of their objective functions is optimal w.r.t approximating the true value function V. Two novel algorithms are proposed to approximate the true value…
Proposes a method to enhance exploration in RL using temporal difference uncertainties.
problem Challenges in estimating uncertainty in non-tabular reinforcement learning settings.
method Estimates uncertainty over value function using temporal difference errors and incorporates it as an intrinsic reward.
result Demonstrates improved exploration in various tasks, including Deep Sea and Atari 2600 environments.
Paper improves TD learning algorithm bounds with linear approx.
problem Sharp bounds for TD method performance in MDPs.
method Polyak-Ruppert averaging, universal step size, refined error bounds, stability of random matrices.
result Near-optimal variance and bias terms achieved.
Paper introduces a new reinforcement learning method with improved performance.
problem Designing and analyzing efficient reinforcement learning algorithms.
method Proximal gradient temporal difference learning (GTD) with accelerated algorithm GTD2-MP.
result GTD algorithms have linear complexity and improved convergence rate.
A new correction term improves sample efficiency in deep reinforcement learning.
problem Momentum accumulation in TD learning leads to doubly stale gradients.
method Proposed a correction term to address the issue of doubly stale gradients.
result Improves sample efficiency in policy evaluation.
Temporal difference learning (TD) is a simple iterative algorithm used to estimate the value function corresponding to a given policy in a Markov decision process. Although TD is one of the most widely used algorithms in reinforcement learning, its theoretical analysis has proved challenging and few guarantees on its s…
Accelerates TD learning for long-horizon reinforcement learning problems.
problem Slow convergence of conventional TD learning in long-horizon tasks.
method Introduces PID Accelerated Temporal Difference (PID TD) learning algorithms.
result Accelerates convergence of TD learning compared to conventional methods.
This paper challenges the conventional wisdom about temporal difference learning's superiority in stitching experience.
problem The effectiveness of temporal difference learning in stitching experience across different reinforcement learning methods.
method Empirical study comparing Monte Carlo and temporal difference learning methods in stitching experience.
result Monte Carlo methods can also achieve experience stitching, challenging the conventional wisdom.
Improved TD learning with neural nets reduces sample complexity and overparameterization.
problem Temporal difference learning with neural networks in large state spaces.
method Projection-free and max-norm regularized Neural TD learning, with Lyapunov drift analysis.
result Max-norm regularization significantly improves TD learning's sample complexity and overparameterization.
Multi-step temporal difference (TD) learning is an important approach in reinforcement learning, as it unifies one-step TD learning with Monte Carlo methods in a way where intermediate algorithms can outperform either extreme. They address a bias-variance trade off between reliance on current estimates, which could be …
Novel bounds improve TD learning consistency in RL.
problem Analyzing Temporal Difference learning's performance.
method High-dimensional concentration inequalities and Berry-Esseen bounds for Markov chain induced martingales.
result Sharp high-probability consistency guarantee for TD learning, matching asymptotic variance up to logarithmic factors.
New TD method stabilizes average-reward learning.
problem Stability issues in average-reward TD learning.
method Implicit fixed point update for average-reward TD(λ). result Improved numerical stability and broader step-size range.
Improved TD learning with tail averaging and regularization achieves optimal convergence rates.
problem Convergence analysis of TD learning with linear function approximation.
method Tail-averaging and regularization applied to TD learning algorithm.
result Achieves optimal O(1/t) convergence rate in expectation and with high probability. Unified framework for PE and TD methods in continuous time and space.
problem Policy evaluation and TD learning in continuous settings.
method Martingale characterization for designing PE algorithms.
result Convergent time-discretized algorithms converge to continuous-time counterparts.
We present the first provably convergent two-timescale off-policy actor-critic algorithm (COF-PAC) with function approximation. Key to COF-PAC is the introduction of a new critic, the emphasis critic, which is trained via Gradient Emphasis Learning (GEM), a novel combination of the key ideas of Gradient Temporal Differ…
New algorithms for collaborative reinforcement learning with limited communication.
problem Efficiently learning value functions in multi-agent systems with strict information constraints.
method Distributed gradient-based temporal difference algorithms with consensus schemes.
result Parameter estimates converge to ODEs with defined invariant sets under general assumptions.
Enhances reinforcement learning uncertainty estimation with a generalized Gaussian error model.
problem Inaccurate error representations and compromised uncertainty estimation in conventional uncertainty-aware TD learning.
method Introduces a novel framework for generalized Gaussian error modeling in deep reinforcement learning, incorporating higher-order moments, particularly kurtosis, to improve uncertainty estimation and mitigation.
result Significant performance gains in policy gradient algorithms with the proposed framework.
TD learning reduces prediction error in Markov chain problems.
problem Estimating value functions in Markov chains with temporal inconsistency.
method Temporal difference learning minimizes temporal inconsistency between successive estimates.
result TD learning can significantly reduce mean-squared error in value estimates.
New algorithms improve reinforcement learning stability and performance.
problem Stability issues in TD learning algorithms with function approximation and off-policy sampling.
method Developed and adapted emphatic temporal difference (ETD(λ)) algorithms for deep reinforcement learning. result Demonstrated improved performance in Atari games and small problems.
TD learning reduces interference, leading to better generalization.
problem Understanding and reducing interference in TD learning for better generalization.
method Analyzing the inner product of gradients as interference, comparing TD and supervised learning, and examining the dynamics of interference and bootstrapping.
result TD learning leads to low-interference, under-generalizing parameters, while supervised learning does the opposite.
TDprop uses Jacobi preconditioning to improve adaptive optimizers in Deep RL.
problem Improving performance of adaptive optimizers in Deep RL.
method TDprop computes per-parameter learning rates based on Jacobi preconditioning of the TD update rule.
result TDprop matches or exceeds Adam's performance in Deep RL experiments, suggesting Jacobi preconditioning can improve adaptive methods.
New algorithms improve distributional TD learning with linear approximations.
problem Estimating return distributions in reinforcement learning.
method Fine-grained analysis of linear-categorical Bellman equation, variance reduction techniques.
result Tight sample complexity bounds for distributional TD learning with linear approximations.
New algorithm for RL using mean embeddings of return distributions.
problem Improving reinforcement learning algorithms for dynamic programming.
method Mean embeddings of return distributions, novel algorithms for RL.
result Asymptotic convergence and improved performance in deep RL.
One of the main obstacles to broad application of reinforcement learning methods is the parameter sensitivity of our core learning algorithms. In many large-scale applications, online computation and function approximation represent key strategies in scaling up reinforcement learning algorithms. In this setting, we hav…
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 statistical methods improve TD learning for policy evaluation.
problem Improving statistical inference for reinforcement learning.
method Polyak-Ruppert averaging, refined high-dimensional Berry-Esseen bounds, online plug-in estimator, asymptotic covariance matrix.
result Guaranteed finite-sample coverage of confidence regions and simultaneous confidence intervals.
We present for the first time an asymptotic convergence analysis of two time-scale stochastic approximation driven by `controlled' Markov noise. In particular, both the faster and slower recursions have non-additive controlled Markov noise components in addition to martingale difference noise. We analyze the asymptotic…
Proposes Deep LTMLE for estimating dynamic treatment effects in longitudinal studies.
problem Estimating counterfactual mean outcomes under dynamic treatment policies in longitudinal settings.
method Uses a transformer architecture with temporal-difference learning for initial estimation, followed by TMLE correction and statistical inference.
result Demonstrates superior performance in complex, long-term scenarios compared to existing methods.
Temporal Difference Learning analysis under non-i.i.d. data and nonlinear approximation.
problem Finite-sample behavior of TD(0) under non-i.i.d. data and nonlinear approximation.
method High-probability, finite-sample analysis of vanilla TD(0) on polynomially mixing Markov data, assuming Holder continuity and bounded generalized gradients.
result Bounds on the convergence rate of TD(0) with high probability, matching known i.i.d. rates and holding even with nonstationary initialization.
Study uses online bootstrap for RL inference, showing effectiveness.
problem Statistical inference for RL parameters in online settings.
method Online bootstrap method applied to TD and GTD algorithms in RL.
result Method is distributionally consistent for policy evaluation inference.
Off-policy reinforcement learning has many applications including: learning from demonstration, learning multiple goal seeking policies in parallel, and representing predictive knowledge. Recently there has been an proliferation of new policy-evaluation algorithms that fill a longstanding algorithmic void in reinforcem…
Introduces TDRC to balance TD's ease and soundness.
problem TD learning's instability and divergence issues.
method Gradient Temporal-Difference Learning with Regularized Corrections (TDRC).
result TDRC performs as well as TD when TD works, but is sound in divergent cases.
PER-ETD improves ETD by reducing variance to polynomial complexity.
problem Large variance in ETD leading to exponential sample complexity.
method Periodically restart and update the follow-on trace for a finite period.
result PER-ETD converges to the same fixed point as ETD but with improved sample complexity.
Recently, a new multi-step temporal learning algorithm, called Q(σ), unifies n-step Tree-Backup (when σ=0) and n-step Sarsa (when σ=1) by introducing a sampling parameter σ. However, similar to other multi-step temporal-difference learning algorithms, Q(σ) needs much memory consumption and computation tim…
ReaPER improves learning efficiency by prioritizing reliable experiences.
problem Inefficient sampling of past experiences in reinforcement learning.
method Introducing a novel measure of reliability to prioritize experiences in PER.
result ReaPER outperforms PER in various environments, including Atari-10.
Paper improves confidence intervals for LSA with multiplier bootstrap.
problem Improving confidence intervals for parameter estimation in LSA.
method Berry-Esseen bound for multivariate normal approximation and multiplier bootstrap.
result Valid confidence intervals for parameter estimation in LSA.
New algorithm optimally evaluates policies with linear approximations.
problem Policy evaluation with linear function approximation.
method Accelerated, variance-reduced fast temporal difference algorithm (VRFTD).
result VRFTD matches both deterministic and stochastic lower bounds.
Unified view on selective credit assignment for reinforcement learning.
problem Efficient credit assignment in reinforcement learning.
method Unified temporal-difference algorithms with selective weightings.
result New algorithms for backward credit assignment and off-policy learning.
To act and plan in complex environments, we posit that agents should have a mental simulator of the world with three characteristics: (a) it should build an abstract state representing the condition of the world; (b) it should form a belief which represents uncertainty on the world; (c) it should go beyond simple step-…
This paper presents four different ways of looking at the well-known Least Squares Temporal Differences (LSTD) algorithm for computing the value function of a Markov Reward Process, each of them leading to different insights: the operator-theory approach via the Galerkin method, the statistical approach via instrumenta…
We introduce a novel Deep Reinforcement Learning (DRL) algorithm called Deep Quality-Value (DQV) Learning. DQV uses temporal-difference learning to train a Value neural network and uses this network for training a second Quality-value network that learns to estimate state-action values. We first test DQV's update rules…