New algorithms improve policy evaluation in reinforcement learning.
problem Off-policy stability and on-policy efficiency issues in policy evaluation.
method Introduced novel algorithms using oblique projection method.
result Demonstrated both off-policy stability and on-policy efficiency.
Memory-efficient algorithm reduces variance in off-policy RL.
problem High variance in off-policy policy optimization.
method Memory-efficient, stochastically variance-reduced algorithm using off-policy samples.
result Empirically validated effectiveness of the proposed algorithm.
Learning the value function of a given policy (target policy) from the data samples obtained from a different policy (behavior policy) is an important problem in Reinforcement Learning (RL). This problem is studied under the setting of off-policy prediction. Temporal Difference (TD) learning algorithms are a popular cl…
We make policy optimization algorithms batch size-invariant by decoupling proximal and behavior policies.
problem Some policy optimization algorithms do not have batch size-invariance, leading to inefficiencies.
method We decouple the proximal policy from the behavior policy to achieve batch size-invariance.
result Our approach makes policy optimization algorithms more efficient and allows them to use stale data more effectively.
The paper interprets policy-gradient algorithms using continuation theory.
problem Optimizing nonconvex functions in reinforcement learning.
method Formulates policy optimization as optimization by continuation, interprets policy-gradient algorithms as implicitly optimizing deterministic policies.
result Exploration in policy-gradient algorithms is seen as computing a continuation of the return of the policy.
We propose a new objective, the counterfactual objective, unifying existing objectives for off-policy policy gradient algorithms in the continuing reinforcement learning (RL) setting. Compared to the commonly used excursion objective, which can be misleading about the performance of the target policy when deployed, our…
Improves reinforcement learning stability and efficiency.
problem Combining stability and efficiency in reinforcement learning.
method Combines on-policy stability with off-policy sample reuse.
result Demonstrates improved performance in both theory and practice.
On-policy reinforcement learning (RL) algorithms have high sample complexity while off-policy algorithms are difficult to tune. Merging the two holds the promise to develop efficient algorithms that generalize across diverse environments. It is however challenging in practice to find suitable hyper-parameters that gove…
Develops ODRPO to improve RL algorithms with better performance and stability.
problem RL algorithms converge to sub-optimal solutions due to limited policy representation.
method Integrates DRO approach to solve trust region constrained optimization problem without parameterizing policies.
result Achieves globally optimal policy update and higher sample efficiency.
We study the problem of off-policy critic evaluation in several variants of value-based off-policy actor-critic algorithms. Off-policy actor-critic algorithms require an off-policy critic evaluation step, to estimate the value of the new policy after every policy gradient update. Despite enormous success of off-policy …
Accelerates policy optimization in RL with optimistic and adaptive updates.
problem Improving policy optimization methods in reinforcement learning.
method Integrates foresight into policy improvement step via optimistic and adaptive updates.
result Designs an optimistic policy gradient algorithm, adaptive via meta-gradient learning.
Optimistic NPG improves policy optimization in online RL with efficient sample complexity.
problem Limited theoretical understanding of policy optimization, especially in online RL.
method Combines natural policy gradient with optimistic policy evaluation.
result Achieves optimal dimension dependence sample complexity for learning near-optimal policies.
Recent successful deep reinforcement learning algorithms, such as Trust Region Policy Optimization (TRPO) or Proximal Policy Optimization (PPO), are fundamentally variations of conservative policy iteration (CPI). These algorithms iterate policy evaluation followed by a softened policy improvement step. As so, they are…
New algorithm for sequential off-policy learning improves performance over batch methods.
problem Training policies from logged interaction data in a sequential setting.
method Combines Logarithmic Smoothing with online PAC-Bayesian tools.
result Improves performance and accelerates convergence in sequential off-policy learning.
STORM-PG uses momentum for faster policy gradient updates.
problem Improving policy gradient methods for reinforcement learning.
method Introduces STORM-PG, a SARAH-based algorithm with exponential moving average.
result Achieves O(1/ε3) sample complexity, matching best-known rate. End-to-end policy learning method improves CATE estimation.
problem Learning optimal treatment policies from partially observed data.
method Modified causal forest for policy learning.
result Maximizing policy value is equivalent to minimizing CATE.
New algorithm finds near-optimal policies efficiently in zero-sum games.
problem Lack of provable efficiency guarantees for policy optimization in zero-sum games.
method Policy optimization algorithm with function approximation.
result Proves efficient convergence to near-optimal policies with polynomial samples and iterations.
The paper tackles robust reinforcement learning with performance guarantees.
problem Finding a robust policy for RMDP with state space uncertainties.
method Proposes RLSPI algorithm for learning optimal robust policy with performance bounds.
result Demonstrates the performance of RLSPI on standard benchmark problems.
Algorithmic stablecoins optimize monetary policy to balance price stability.
problem Persistent inflation from centralized monetary policy.
method Propose and study a rule-based monetary policy model for algorithmic stablecoins.
result Optimal trade-off between price stability and supply stability.
PPO algorithm converges to global optimality in multi-agent reinforcement learning.
problem Designing statistical guarantees for policy optimization methods in multi-agent reinforcement learning.
method Leveraging a multi-agent performance difference lemma, a localized action value function is used as a descent direction for each local policy, leading to a multi-agent PPO algorithm.
result The multi-agent PPO algorithm converges to the globally optimal policy at a sublinear rate under standard regularity conditions.
GPE algorithm optimizes nonparametric contextual bandits with efficient regret bounds.
problem Optimizing nonparametric contextual bandits with efficient regret bounds.
method Inspired by Policy Elimination, GPE uses oracle-efficient techniques for nonparametric classes with infinite VC-dimension.
result GPE is regret-optimal for policy classes with integrable entropy, and for larger entropy, it provides an ε-greedy algorithm with matching regret bounds. Study shows code-level optimizations significantly impact deep RL algorithms.
problem Understanding the impact of implementation details on deep RL algorithms.
method Case study on PPO and TRPO, investigating the effects of code-level optimizations.
result Code-level optimizations are crucial for performance in deep RL algorithms.
New RL algorithms learn policies competitive with best in class without assuming optimal policy.
problem Agnostic RL where best policy not necessarily optimal.
method First-order optimization in non-Euclidean space, reducing to policy learning.
result Sample complexity upper bounds for three algorithms under VGD condition.
Off-policy deep reinforcement learning (RL) algorithms are incapable of learning solely from batch offline data without online interactions with the environment, due to the phenomenon known as \textit{extrapolation error}. This is often due to past data available in the replay buffer that may be quite different from th…
PBVFs generalize across policies using learned value functions.
problem RL algorithms forget information about old policies when updating value functions to track the learned policy.
method Introduce Parameter-Based Value Functions (PBVFs) that include policy parameters in their inputs, enabling them to generalize across different policies.
result PBVFs enable zero-shot learning of new policies that outperform any policy seen during training.
New algorithm stabilizes RL policy learning through divergence regularization.
problem Stabilize policy learning and improve performance in RL.
method Proximity term constraining discounted state-action visitation distributions to be close to each other.
result Proposed algorithm improves stability and final performance in RL tasks.
Learning policies on data synthesized by models can in principle quench the thirst of reinforcement learning algorithms for large amounts of real experience, which is often costly to acquire. However, simulating plausible experience de novo is a hard problem for many complex environments, often resulting in biases for …
One-step policy improvement outperforms iterative RL methods on D4RL.
problem Improving offline RL without off-policy evaluation.
method One-step constrained/regularized policy improvement using on-policy Q estimates.
result One-step algorithm outperforms iterative algorithms on D4RL benchmark.
We study continuous action reinforcement learning problems in which it is crucial that the agent interacts with the environment only through safe policies, i.e.,~policies that do not take the agent to undesirable situations. We formulate these problems as constrained Markov decision processes (CMDPs) and present safe p…
To overcome the curses of dimensionality and modeling of Dynamic Programming (DP) methods to solve Markov Decision Process (MDP) problems, Reinforcement Learning (RL) methods are adopted in practice. Contrary to traditional RL algorithms which do not consider the structural properties of the optimal policy, we propose …
Study improves off-policy evaluation from non-i.i.d. bandit samples.
problem Improving off-policy evaluation from non-independent bandit samples.
method Constructing an estimator from a standardized martingale difference sequence.
result Proposed estimator performs better than existing methods.
This paper bridges offline and online RL by studying policy finetuning with a reference policy.
problem Sample-efficient reinforcement learning in online and offline settings.
method Design of policy finetuning algorithms and analysis of sample complexity.
result Theoretical analysis shows that the optimal policy finetuning algorithm is either offline reduction or purely online RL.
New RL algorithms improve control tasks with data reuse.
problem Real-world control requires performance guarantees and data efficiency.
method Generalized Policy Improvement combining on-policy guarantees and sample reuse.
result Extensive experimental analysis shows benefits of new algorithms.
In this paper, a new population-guided parallel learning scheme is proposed to enhance the performance of off-policy reinforcement learning (RL). In the proposed scheme, multiple identical learners with their own value-functions and policies share a common experience replay buffer, and search a good policy in collabora…
Improves RL algorithms with two techniques.
problem Enhance off-policy RL performance.
method Formulates RL as proximal point iteration; uses value functions for improved action value estimate.
result Significant performance improvement on RL benchmarks.
Algorithm improves reinforcement learning policies using offline data.
problem Improving reinforcement learning policies with limited online data.
method Designs a single non-reactive policy using offline data with provable guarantees.
result Algorithm achieves better policy quality with less online data.
Algorithm safely learns from sub-optimal baseline policies while satisfying constraints.
problem Safe reinforcement learning with constraints when baseline policy is sub-optimal.
method Iterative policy optimization alternating between return maximization, baseline distance minimization, and constraint projection.
result Consistently outperforms baselines, achieving 10x fewer constraint violations and 40% higher reward.
We develop an off-policy actor-critic algorithm for learning an optimal policy from a training set composed of data from multiple individuals. This algorithm is developed with a view towards its use in mobile health.
This paper extends off-policy reinforcement learning to the multi-agent case in which a set of networked agents communicating with their neighbors according to a time-varying graph collaboratively evaluates and improves a target policy while following a distinct behavior policy. To this end, the paper develops a multi-…
Entropy regularization improves policy optimization in reinforcement learning.
problem Improving policy optimization in reinforcement learning.
method Entropy regularization is introduced to soften the greedy policy towards a more diverse softmax policy, leading to a continuously parameterized algorithm that interpolates between policy gradient and Q-learning.
result An intermediate algorithm can improve performance in reinforcement learning.
Boosting for off-policy learning reduces empirical risk.
problem Learning from logged bandit feedback without labeled data.
method A boosting algorithm optimizing policy's expected reward.
result Excess empirical risk decreases with each round of boosting.
Theoretical analysis confirms non-conservative algorithms can converge to optimal policies.
problem Theoretical guarantees for non-conservative reinforcement learning algorithms.
method Theoretical analysis of Peng's Q(λ) algorithm. result Peng's Q(λ) converges to an optimal policy under certain conditions. Off-policy learning in dynamic decision problems is essential for providing strong evidence that a new policy is better than the one in use. But how can we prove superiority without testing the new policy? To answer this question, we introduce the G-SCOPE algorithm that evaluates a new policy based on data generated by…
Imitation learning can reproduce policies by observing experts, which poses a problem regarding policy privacy. Policies, such as human, or policies on deployed robots, can all be cloned without consent from the owners. How can we protect against external observers cloning our proprietary policies? To answer this quest…
In this work, we take a fresh look at some old and new algorithms for off-policy, return-based reinforcement learning. Expressing these in a common form, we derive a novel algorithm, Retrace(λ), with three desired properties: (1) it has low variance; (2) it safely uses samples collected from any behaviour policy, wha…
Building upon the recent success of deep reinforcement learning methods, we investigate the possibility of on-policy reinforcement learning improvement by reusing the data from several consecutive policies. On-policy methods bring many benefits, such as ability to evaluate each resulting policy. However, they usually d…
Mirror descent method improved RL algorithms.
problem Improving RL algorithms for better performance.
method Mirror descent method applied to RL, solving trust-region problems.
result MDPO outperforms or matches other RL algorithms in continuous control tasks.
Optimizes Thompson sampling policies using policy gradient methods.
problem Improving Thompson sampling in bandit problems.
method Applies policy gradient algorithms to optimize Thompson sampling policies.
result Direct policy search on Thompson sampling improves performance.