AE-DQN learns to eliminate sub-optimal actions in complex RL environments.
problem Challenges in RL with many actions, especially redundancy.
method Combines DQN with AEN that predicts invalid actions.
result Significant speedup and robustness in games with many actions.
A method for training autonomous vehicles using continuous human feedback to avoid sub-optimal decisions.
problem Training autonomous vehicles with limited and potentially sub-optimal human demonstrations.
method Continuous scalar feedback for each action to learn from sub-optimal demonstrations and evaluative feedback.
result The proposed method outperforms supervised learning on positive examples alone and learns from sub-optimal demonstrations.
New algorithm reduces sample complexity for planning in MDPs.
problem Planning in MDPs with unknown transitions.
method MDP-GapE, a trajectory-based MCTS algorithm.
result Proves upper bound on sample complexity in terms of sub-optimality gaps.
Decouples critic chunk length from policy to improve policy reactivity and performance.
problem Bootstrapping bias and difficulty in extracting optimal policies from chunked critics.
method Optimizes policy against a distilled critic for partial action chunks, allowing shorter chunks for policy.
result Reliably outperforms prior methods on long-horizon offline goal-conditioned tasks.
LEARN-SAM improves RL from sub-optimal demonstrations by localizing expert policies and selectively using demonstrations.
problem Improving RL from sub-optimal or sparse demonstrations.
method Local Ensemble and Reparameterization with Split and Merge of expert policies (LEARN-SAM).
result LEARN-SAM boosts learning speed and accuracy by selectively using demonstrations.
We study an original problem of pure exploration in a strategic bandit model motivated by Monte Carlo Tree Search. It consists in identifying the best action in a game, when the player may sample random outcomes of sequentially chosen pairs of actions. We propose two strategies for the fixed-confidence setting: Maximin…
Logarithmic regret achieved in Q-learning with positive gap.
problem Achieving logarithmic cumulative regret in Q-learning with positive sub-optimality gap.
method Optimistic Q-learning with logarithmic regret bound.
result Logarithmic cumulative regret bound proven for optimistic Q-learning.
We study the sparse entropy-regularized reinforcement learning (ERL) problem in which the entropy term is a special form of the Tsallis entropy. The optimal policy of this formulation is sparse, i.e.,~at each state, it has non-zero probability for only a small number of actions. This addresses the main drawback of the …
Approximate models help RL by reducing policy search space.
problem How much does an approximate model help in learning near-optimal policies in RL?
method Study sample complexity in RL with an approximate model, providing an algorithm and a lower bound.
result An approximate model can reduce sample complexity by eliminating sub-optimal actions.
UTE improves reinforcement learning by measuring action uncertainty, enhancing policy learning efficiency.
problem Degrading performance of action repetition in reinforcement learning, especially with sub-optimal actions.
method UTE uses ensemble methods to measure uncertainty during action extension, allowing strategic exploration or certainty.
result UTE outperforms existing action repetition algorithms, significantly enhancing policy learning efficiency.
Paper tackles sample-efficient RL for linearly realizable MDPs with limited revisiting.
problem Sample-efficient reinforcement learning for linearly realizable MDPs with limited revisiting.
method Develops a new sampling protocol that allows for backtracking and revisiting states in a controlled manner.
result Achieves polynomial sample complexity scaling with feature dimension, horizon, and inverse sub-optimality gap.
Paper introduces a new value function for state transitions and optimal policy learning.
problem Learning optimal policies from state transitions and actions.
method Develops a forward dynamics model to maximize a novel value function Q ( s , s ′ ) Q(s, s') Q ( s , s ′ ) . result Demonstrates benefits in value function transfer, redundant action spaces, and off-policy learning.
We address reinforcement learning problems with finite state and action spaces where the underlying MDP has some known structure that could be potentially exploited to minimize the exploration rates of suboptimal (state, action) pairs. For any arbitrary structure, we derive problem-specific regret lower bounds satisfie…
The Prescriptive Canvas improves business outcomes by directly prescribing actions based on predictions.
problem Sub-optimal performance in business projects due to a two-step approach of prediction and decision-making.
method The Prescriptive Canvas methodology for framing and communicating actions directly based on predictions.
result Improves framing and communication across stakeholders for successful business impact.
Enhances reinforcement learning safety through risk-averse exploration.
problem Safety concerns in reinforcement learning due to sub-optimal actions.
method Distributionally robust policy iteration scheme with lower bound guarantees.
result Efficient algorithm that prevents poor decisions and converges to optimal policy.
This study examines how learning algorithms affect collective action in machine learning.
problem The impact of collective action on machine learning is limited when not considering the choice of learning algorithms.
method Focuses on distributionally robust optimization and stochastic gradient descent, analyzing their effects on collective success.
result The choice of learning algorithm significantly impacts the effective size and success of a collective in machine learning.
New algorithm tackles nonstationary linear bandits with latent dynamics.
problem Nonstationary bandit problem with latent states and unknown dynamics.
method Explore-then-commit algorithm with exploration and commitment phases.
result Achieves i l d e O ( T 2 / 3 ) ilde{\mathcal{O}}(T^{2/3}) i l d e O ( T 2/3 ) regret. A new method for continuous control avoids local movement issues.
problem Limitations of policy gradient methods in continuous control.
method Distributional framework and Generative Actor Critic (GAC) method.
result GAC outperforms policy gradient methods in continuous domains.
Improves BC policies by generating new plausible trajectories.
problem Sub-optimal data quality in BC leads to poor policy performance.
method Trajectory Stitching (TS) generates new plausible transitions.
result TS significantly improves behavioural policies over original data.
Reinforcement learning aids decision-making in economics and finance.
problem Optimal decision-making in dynamic, uncertain environments.
method Reinforcement learning algorithms to learn optimal policies.
result Deep learning enhances solving complex behavioral problems.
Logarithmic regret achieved in RL with linear function approximation.
problem Achieving logarithmic regret in reinforcement learning with linear function approximation.
method LSVI-UCB for linear MDP assumption, UCRL-VTR for linear mixture MDP assumption.
result Logarithmic regret bounds established for RL with linear function approximation.
Unified hybrid RL algorithm improves online RL performance with offline data.
problem Improving reinforcement learning performance with limited online data.
method A unified hybrid RL algorithm combining offline and online data.
result Unified algorithm achieves state-of-the-art results in sub-optimality gap and online learning regret.
Proposes Constrained Q-learning for reinforcement learning with constraints.
problem Optimizing multiple objectives while adhering to constraints in reinforcement learning.
method Directly restricts the action space in Q-update to learn optimal Q-function for constrained MDP.
result Improves safety and optimality in high-level decision making for autonomous driving.
Certainty equivalent controllers perform nearly optimally in LQ control problems with unknown dynamics.
problem Optimizing control in systems with unknown transition dynamics.
method Analysis of certainty equivalent controllers and comparison to optimal LQ controllers, using perturbation bounds for discrete Riccati equations.
result Sub-optimality gap scales as the square of the parameter error, improving upon previous results.
Improves RL from historical data by stitching trajectories.
problem Lack of high-quality data for offline RL.
method Trajectory Stitching (TS) to augment historical data with synthetic actions.
result Improves RL policy performance over baseline.
Paper bounds PAC RL sample complexity in deterministic MDPs.
problem Identify ε-optimal policy with high probability.
method Proposes nearly matching upper and lower bounds on sample complexity, introduces deterministic return gap, uses graph-theoretical concepts and maximum-coverage exploration.
result First nearly matching upper and lower bounds on sample complexity for PAC RL in deterministic MDPs.
Paper tackles covariate shift in offline IL using less proficient behavior data.
problem Mitigating covariate shift in Imitation Learning with limited data coverage.
method Model-based IL from Offline Data (MILO) framework.
result MILO can combat covariate shift even with sub-optimal behavior policy data.
Q-learning requires more samples than minimax bounds suggest for optimal Q-function approximation.
problem Understanding the sample complexity of Q-learning in synchronous settings.
method Analyzing Q-learning in synchronous MDPs with state and action spaces, proving minimax optimal sample complexity for TD learning and Q-learning under certain conditions.
result Q-learning requires more samples than minimax bounds suggest, revealing strict sub-optimality when action space is more than one.
SAIL learns from sub-optimal demonstrations to improve sample efficiency in sparse reward tasks.
problem Reducing sample complexity in sparse-rewarded tasks.
method Self-Adaptive Imitation Learning (SAIL) that exploits sub-optimal demonstrations and efficient exploration.
result Significantly improved sample efficiency and better final performance across various tasks.
New method achieves near-optimal regret without simulator.
problem Adversarial linear contextual bandits with unknown loss vectors.
method Near-optimally reduces regret to sqrt(T) without simulator.
result Achieves regret of sqrt(T) without simulator, improving existing methods.
New measure quantifies function similarity for optimization.
problem Measuring similarity between functions for optimization.
method Quantifies sub-optimality gaps and operation rules.
result Unified measure for various functional similarities.
New proof shows deep neural nets can have sub-optimal local minima.
problem Can over-parameterization eliminate sub-optimal local minima in deep neural networks?
method Counter-example with generic input data and non-linear activation functions.
result Sub-optimal local minima exist in deep neural networks regardless of width.
Unified approach translates classic bandit algorithms to structured settings.
problem Finite-armed structured bandit problem with unknown reward functions.
method Gradual estimation of hidden parameter θ* and use in mean reward functions.
result Structured bandit versions of UCB achieve bounded regret in practical scenarios.
Algorithm reduces regret in misspecified linear contextual bandits.
problem Misspecified linear contextual bandits with bounded misspecification.
method Data selection scheme for online regression, leveraging uncertainty.
result Regret bound of O ~ ( d 2 / Δ ) \tilde O(d^2/Δ) O ~ ( d 2 /Δ ) when ζ ≤ O ~ ( Δ / d ) ζ \leq \tilde O(Δ/\sqrt{d}) ζ ≤ O ~ ( Δ/ d ) . Wide networks avoid bad local minima, proving phase transition.
problem Understanding the optimization landscape of neural networks.
method Proving phase transitions in loss surfaces of wide and narrow networks.
result Phase transition from narrow to wide networks: no bad local minima.
A new algorithm trains experts to safely guide agents in partially observed environments.
problem Existing imitation learning methods for POMDPs can lead to sub-optimal or unsafe policies.
method Derive an objective to encourage the expert to maximize the agent's reward, then use it to train both expert and agent.
result The algorithm produces an expert policy that the agent can safely imitate, outperforming fixed expert policies.
The paper tackles combinatorial pure exploration with various feedback structures and proposes efficient algorithms.
problem Identifying the optimal action in a combinatorial space with limited feedback and nonlinear rewards.
method Designs polynomial-time adaptive algorithms for CPE-BL and CPE-PL, providing sample complexity analyses.
result The proposed algorithms achieve sample complexity close to lower bounds and outperform existing methods.
Study on individual regret in cooperative MAB with agents communicating over a graph.
problem Individual regret in cooperative stochastic multi-armed bandits with communication constraints.
method Analyzed COOP-SE algorithm, derived individual regret bounds under various communication constraints.
result First to show an individual regret bound in cooperative stochastic MAB independent of graph diameter.
New algorithm optimizes multi-objective outcomes in uncertain environments.
problem Optimizing global concave rewards in online Markov decision processes with multiple actions.
method No-regret algorithm based on online convex optimization and UCRL2, with a gradient threshold procedure.
result Non-stationary policy diversifies outcomes to optimize the global concave reward.
TRAiL is a linear bandit algorithm that ensures optimal regret and guarantees inference quality.
problem Optimal regret and inference quality in linear bandits with convex action sets.
method TRAiL estimates the parameter through regularized least squares and perturbs the action set along the tangent plane.
result TRAiL achieves an Ω ( T ) Ω(\sqrt{T}) Ω ( T ) upper bound on cumulative regret with high probability. Policy gradient converges linearly with Hadamard parameterization in tabular settings.
problem Convergence of policy gradient methods under Hadamard parameterization.
method Studied convergence rate and established linear convergence after k 0 k_0 k 0 iterations. result Algorithm converges linearly with rate $O( rac{1}{k})$ and faster locally after k 0 k_0 k 0 . New algorithm tackles multi-agent reinforcement learning with optimal convergence rate.
problem Multi-agent reinforcement learning with large state spaces and linear function approximations.
method Refined AVLPR framework with data-dependent pessimistic estimation and action-dependent bonuses.
result First algorithm with optimal O ( T − 1 / 2 ) O(T^{-1/2}) O ( T − 1/2 ) convergence rate and no poly( A max A_{\max} A m a x ) dependency. Aims to create safe reinforcement learning policies by considering individual harm.
problem Optimal policies for a population may harm certain individuals.
method Formalizes individual harm, proposes a two-stage procedure, and establishes finite-sample properties.
result Learned policies maximize expected return while minimizing harm.
New algorithm achieves optimal regret in average reward MDPs without prior bias information.
problem Achieving optimal regret in average reward MDPs with computational efficiency and without prior bias information.
method Projective Mitigated Extended Value Iteration (PMEVI) to compute bias-constrained optimal policies efficiently.
result First tractable algorithm with minimax optimal regret of O ~ ( s p ( h ∗ ) S A T ) \widetilde{\mathrm{O}}(\sqrt{\mathrm{sp}(h^*) S A T}) O ( sp ( h ∗ ) S A T ) . New algorithm reduces sample complexity for new tasks by leveraging prior knowledge.
problem Designing reinforcement learning agents that reduce sample complexity for new tasks.
method Designing an algorithm that quickly identifies an accurate solution by seeking informative state-action pairs from related tasks, using a generative model.
result PAC bounds on sample complexity demonstrate the benefits of using prior knowledge.
Researchers develop a method to control nonlinear systems with Koopman operator regression.
problem Controlling nonlinear systems with finite action spaces.
method Koopman operator regression for dynamics estimation and model predictive control for control.
result The method yields a linear switching predictive model for control.
Softmax Bellman operator improves Q-function performance in RL despite sub-optimality.
problem Softmax Bellman operator's impact on value functions in RL is problematic.
method Revisited theoretical properties of softmax Bellman operator, proving convergence and overestimation reduction.
result Softmax Bellman operator leads to superior policies in practice, even outperforming double Q-learning.
Study online learning with feedback graphs and switching costs, providing algorithms and optimal regret bounds.
problem Online learning with partial feedback and switching costs.
method Analysis of feedback graphs, lower bound on expected regret, new algorithms (Threshold Based EXP3, EXP3. SC).
result Order optimal algorithms for specific cases and Threshold Based EXP3 outperforms in empirical evaluations.