EBU improves deep RL with episodic backward updates, achieving human-level performance with less data.
problem Efficiently learn from sparse and delayed rewards in reinforcement learning.
method Episodic Backward Update (EBU) algorithm that recursively propagates value updates.
result EBU achieves human-level performance in Atari games with significantly fewer samples.
This work improves policy-based training by proposing an evaluation balance objective for GFlowNets.
problem Reliable estimation of policy divergence under directed acyclic graphs remains challenging.
method Proposes an evaluation balance objective over partial episodes to measure policy divergence and improve policy-based training reliability.
result Evaluation balance strengthens policy-based training reliability and broadens its flexibility.
We investigate different strategies for active learning with Bayesian deep neural networks. We focus our analysis on scenarios where new, unlabeled data is obtained episodically, such as commonly encountered in mobile robotics applications. An evaluation of different strategies for acquisition, updating, and final trai…
The study examines backward compatibility issues in ML systems, especially with noisy data.
problem Backward compatibility challenges in ML systems, especially with noisy data.
method Empirical analysis of ML systems across different architectures and datasets, focusing on data shifts and noise.
result Backward compatibility issues arise even without data shift due to optimization stochasticity and training on large-scale noisy datasets can significantly decrease compatibility.
BC-Aligner maintains backward compatibility of embeddings after frequent updates.
problem Updating embeddings without requiring consumer teams to retrain their models.
method Learning backward compatible embeddings through BC-Aligner.
result BC-Aligner maintains backward compatibility with existing unintended tasks after multiple model version updates.
Algorithm improves online learning in adversarial bandits.
problem Online learning in adversarial multi-armed bandits with non-uniform best arm distribution.
method Online-within-online setup, inner and outer learners, leveraging non-uniform empirical distribution of best arms.
result Improves regret bounds for non-uniform best arm distributions.
New model for personalized online advertising with multi-user interaction.
problem Realistic online advertising scenarios with multiple users interacting simultaneously.
method Introduces Multi-User Contextual Cascading Bandit (MCCB) model and proposes UCBBP and AUCBBP algorithms.
result Proves UCBBP and AUCBBP achieve optimal regret bounds for multi-user context.
The study learns neural update rules by remembering past experiences.
problem Developing efficient online learning rules for neural networks.
method Representing neurons with vectors, using meta-neural networks for updates, and training for remembering past experiences.
result The approach reveals insights into learning rules and could be used for complex tasks like episodic memory.
Most provably-efficient learning algorithms introduce optimism about poorly-understood states and actions to encourage exploration. We study an alternative approach for efficient exploration, posterior sampling for reinforcement learning (PSRL). This algorithm proceeds in repeated episodes of known duration. At the sta…
Optimistic PPO variant solves linear MDPs with improved regret bound.
problem Understanding theoretical limits of PPO in linear MDPs.
method Proposes an optimistic variant of PPO for episodic adversarial linear MDPs with full-information feedback.
result Establishes a i l d e O ( d 3 / 4 H 2 K 3 / 4 ) ilde{\mathcal{O}}(d^{3/4}H^2K^{3/4}) i l d e O ( d 3/4 H 2 K 3/4 ) regret bound. Deep reinforcement learning methods attain super-human performance in a wide range of environments. Such methods are grossly inefficient, often taking orders of magnitudes more data than humans to achieve reasonable performance. We propose Neural Episodic Control: a deep reinforcement learning agent that is able to rap…
New method improves continual learning by anchoring past knowledge.
problem Catastrophic forgetting in continual learning.
method Bilevel optimization to update current task knowledge while keeping past task predictions.
result Improves accuracy and forgetting metrics compared to experience replay.
New schemes improve lifelong learning by balancing old and new tasks.
problem Catastrophic forgetting in deep neural networks when learning multiple tasks.
method Unified optimization perspective of episodic memory based approaches, introducing MEGA-I and MEGA-II schemes.
result Significant improvement in lifelong learning benchmarks, reducing error by up to 18%.
AMUSE uses reinforcement learning to predict optimal model updates.
problem Concept drift weakens model performance over time.
method Reinforcement learning in a simulated environment.
result AMUSE proactively recommends updates based on performance improvements.
This paper formalises the problem of online algorithm selection in the context of Reinforcement Learning. The setup is as follows: given an episodic task and a finite number of off-policy RL algorithms, a meta-algorithm has to decide which RL algorithm is in control during the next episode so as to maximize the expecte…
No-regret algorithm for contextual RL with GLM mappings.
problem Learning near-optimal policies in episodic MDPs with contextual information.
method Proposes no-regret online RL algorithm using optimistic and randomized exploration methods.
result Improves previous bounds and provides a lower bound for the setting.
UDRL fails to converge in stochastic environments with episodic resets.
problem UDRL's convergence in stochastic environments with resets is questioned.
method UDRL is a supervised learning approach that does not use value functions.
result UDRL diverges in a simple stochastic environment with resets.
New RL algorithm achieves nearly optimal performance for linear MDPs.
problem Optimal reinforcement learning for episodic linear MDPs.
method Weighted linear regression with variance estimator and rare-switching policy.
result Achieves nearly minimax optimal regret i l d e O ( d H 3 K ) ilde O(d\sqrt{H^3K}) i l d e O ( d H 3 K ) . Study shows how to control jump-diffusion processes with stable feedback controls in reinforcement learning.
problem Control jump-diffusion processes with unknown coefficients in reinforcement learning.
method Lipschitz continuous optimal feedback controls, stability analysis of forward-backward SDEs, least-squares algorithm.
result Achieves O ( N ln N ) O(\sqrt{N\ln N}) O ( N ln N ) regret for linear-convex learning problems with jumps. New method shows how order of gradient updates impacts stability and convergence in deep learning.
problem Training deep learning models can be unstable and computationally expensive.
method Theoretical analysis and experiments with backward-SGD.
result The order of gradient updates affects stability and convergence, leading to improved performance.
Gradient-based methods for games suffer from discrete update steps that cause drift, affecting performance.
problem Gradient-based methods for two-player games suffer from drift due to discrete update steps.
method Derived modified continuous dynamical systems to closely follow the discrete dynamics of games.
result Identified distinct components of discretization drift that can alter or destabilize game performance.
SDPA is shown to be an optimal transport problem in deep learning.
problem The mathematical foundation and optimization perspective of SDPA.
method SDPA is shown to be the exact solution to a degenerate, one-sided Entropic Optimal Transport (EOT) problem.
result The SDPA mechanism is a principled mechanism where the forward pass performs optimal inference and the backward pass implements a rational, manifold-aware learning update.
YOPO reduces adversarial training time by one forward/backward pass per group of updates.
problem Adversarial training's high computational cost limits deep learning's robustness.
method YOPO uses the Pontryagin's Maximal Principle to limit propagation to the first layer.
result YOPO achieves comparable defense accuracy with significantly less computational time.
Generalising the idea of the classical EM algorithm that is widely used for computing maximum likelihood estimates, we propose an EM-Control (EM-C) algorithm for solving multi-period finite time horizon stochastic control problems. The new algorithm sequentially updates the control policies in each time period using Mo…
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.
Backpropagation algorithm is indispensable for the training of feedforward neural networks. It requires propagating error gradients sequentially from the output layer all the way back to the input layer. The backward locking in backpropagation algorithm constrains us from updating network layers in parallel and fully l…
AI updates can disrupt human-AI teams; new objective improves compatibility.
problem AI updates can disrupt human-AI team performance.
method Introduce compatibility concept, propose re-training objective.
result Current machine learning algorithms do not produce compatible updates.
New method for dynamic valuation in markets with random endowments.
problem Dynamic valuation in markets with random endowments.
method Developed new FBSDE systems and established optimality conditions.
result Established necessary and sufficient conditions for optimality.
La-MAML improves fast online continual learning with a look-ahead approach.
problem Fast online continual learning with limited model capacity.
method Optimisation-based meta-learning with look-ahead and episodic memory.
result Superior performance on visual classification benchmarks.
New strategy improves liquidity takers' performance in markets with latency.
problem Latency affects liquidity takers' ability to execute limit orders effectively.
method Modelled LOB and MLOs as a marked point process, used variational analysis and FBSDEs to find optimal price limits.
result Optimal trading strategy improves marksmanship in markets with latency.
Paper analyzes distributional reinforcement learning with value function approximation, introducing Bellman unbiasedness and a new algorithm.
problem Improving reinforcement learning by capturing environmental stochasticity and addressing infinite dimensionality.
method Introduces Bellman unbiasedness and proposes SF-LSVI algorithm for provably efficient distributional reinforcement learning.
result Achieves a tight regret bound of O(d_E H^3/2 √K) for distributional reinforcement learning.
New algorithm for Q-learning in MDPs with infinite states and actions.
problem Efficient learning in MDPs with large discrete states and actions.
method Proposes a novel algorithm for MDPs with Lipschitz continuous payoff and transition.
result Achieves nearly optimal regret i l d e O ( K d + 1 d + 2 H 3 ) ilde{\mathcal{O}}(K^{\frac{d+1}{d+2}}\sqrt{H^3}) i l d e O ( K d + 2 d + 1 H 3 ) . Study preference-based reinforcement learning in episodic kernel MDPs.
problem Learning from episodic human preferences in reinforcement learning.
method Developed preference-based value estimation and confidence sets for kernel-based MDPs.
result Proved high-probability regret bounds that converge to optimal policy value.
Tensor trains simplify solving complex PDEs efficiently.
problem Solving high-dimensional parabolic PDEs using traditional methods is computationally infeasible.
method Reformulate PDEs as backward stochastic differential equations and use tensor train format for compression and efficient computation.
result Tensor train methods achieve a good balance between accuracy and computational efficiency.
Algorithm tackles constrained reinforcement learning with concave-convex and knapsack constraints.
problem Constrained episodic reinforcement learning with concave rewards and convex constraints.
method Modular analysis with strong theoretical guarantees for concave-convex and knapsack settings.
result Significantly outperforms existing approaches in constrained episodic environments.
The paper addresses Dyna-style RL's value hallucination issue by proposing a new algorithm.
problem Value hallucination in Dyna-style RL due to bootstrapping simulated states.
method Introduces a new Dyna algorithm using predecessor models with multi-step updates.
result Evidence supports the Hallucinated Value Hypothesis (HVH), suggesting predecessor models with multi-step updates are promising.
Algorithm improves multi-armed bandit performance by transferring reward samples.
problem Sequential multi-armed bandit problem with changing reward distributions.
method UCB algorithm with reward sample transfer.
result Significant improvement in cumulative regret over standard UCB.
A new method for risk-averse decision-making in Markov processes with improved regret bounds.
problem Risk-averse decision-making in Markov processes.
method Introduces mini-batch measures and multipattern risk-averse problems in a feature-based Q Q Q -learning method. result Proves a high-probability regret bound of O ( H 2 N H K ) \mathcal{O}\big(H^2 N^H \sqrt{ K}\big) O ( H 2 N H K ) for the Q Q Q -learning method. BerlinUCB learns from episodic rewards in nonstationary contexts.
problem Online learning with episodic rewards in nonstationary environments.
method BerlinUCB integrates clustering for self-supervision.
result BerlinUCB outperforms standard contextual bandits in various scenarios.
Bio-inspired neural networks use predictive coding for efficient weight updates.
problem Training artificial neural networks efficiently and biologically plausibly.
method Predictive Coding (PC) updates weights locally using only local information.
result PC provides theoretical advantages like automatic gradient scaling.
Backplay improves RL efficiency with demonstrations.
problem Improving sample efficiency in sparse reward environments.
method Constructs a curriculum using a single demonstration to start training near the end of the demo.
result Backplay outperforms other methods in grid worlds and a complex game.
Improved regret bound for online learning in unknown MDPs.
problem Online learning in unknown episodic MDPs with changing loss functions.
method Adapts adversarial MDP model to convex performance criteria using entropic regularization.
result Achieved i l d e O ( L ∣ X ∣ ∣ A ∣ T ) ilde{O}(L|X|\sqrt{|A|T}) i l d e O ( L ∣ X ∣ ∣ A ∣ T ) regret bound. Algorithm reduces episode count for CMDPs with constraints.
problem Online decision-making with constraints in episodic CMDPs.
method Optimistic planning using linear programming for PAC guarantee.
result Probably approximately correct (PAC) guarantee on episode count.
Many interesting real world domains involve reinforcement learning (RL) in partially observable environments. Efficient learning in such domains is important, but existing sample complexity bounds for partially observable RL are at least exponential in the episode length. We give, to our knowledge, the first partially …
New FPI layers enable efficient backpropagation in deep networks.
problem Designing deep neural networks to handle complex constraints.
method Fixed-point iteration layers for forward and backward propagation.
result Backward FPI layer simplifies gradient calculation without explicit Jacobian.
New algorithm learns reward signals from episodic returns for better reinforcement learning.
problem Difficulty in designing reward functions for real-world reinforcement learning tasks.
method Introduces a new algorithm that decomposes episodic returns into time-step rewards using deep neural networks.
result Learning reward signals from episodic returns improves reinforcement learning efficiency.
Q-learning achieves optimal regret with UCB exploration.
problem Proving sample efficiency of model-free RL algorithms.
method Q-learning with UCB exploration, proving regret bound.
result Q-learning achieves optimal regret bound of i l d e O ( H 3 S A T ) ilde{O}(\sqrt{H^3 SAT}) i l d e O ( H 3 S A T ) . Novel method for SDE calibration from sparse data using neural flows.
problem Calibrating SDEs from sparse, noisy observations.
method Characterization of posterior SDE using neural networks trained to solve a PDE with multiplicative updates.
result Significant improvement in scalability and accuracy compared to classical methods.