This paper tackles hard exploration in the game Pommerman, improving RL learning.
problem Hard exploration in sparse, delayed, and deceptive reward domains.
method Developed a model-based automatic reasoning module to prune unsafe actions.
result Model-based approach significantly improves RL learning in Pommerman.
A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, whic…
Study reward-free RL in non-linear settings, improving efficiency and removing assumptions.
problem Improving sample efficiency in reward-free reinforcement learning for non-linear function approximation.
method Proposed RFOLIVE algorithm for minimal structural assumptions, analyzed hardness results for reward-free and reward-aware exploration.
result Statistical efficiency and hardness results under various structural assumptions, no need for reachability or explorability assumptions.
BYOL-Explore learns to explore visually-rich environments by predicting world dynamics.
problem Exploration in visually complex environments.
method Optimizes a single prediction loss in latent space to learn world representation, dynamics, and exploration policy.
result Achieves superhuman performance on Atari games with simpler design.
APT-Gen generates tasks to help RL learn in hard problems.
problem Learning in hard exploration problems.
method APT-Gen uses a task generator to create tasks from a parameterized space, balancing performance and similarity to target tasks.
result APT-Gen outperforms baselines in grid world and robotic manipulation tasks.
RECODE uses clustering and embedding to track state visitation counts in RL.
problem Efficient novelty-based exploration in nonstationary RL environments.
method Non-parametric clustering, online density estimation, inverse dynamics loss.
result RECODE achieves state-of-the-art performance in challenging RL tasks.
The paper improves regret lower bounds for communicating MDPs.
problem Regret lower bounds for communicating MDPs.
method Lower bound proof and optimization problem formulation.
result Regret lower bound becomes significantly more complex in communicating MDPs.
Agent learns directed exploration policies to improve performance in hard games.
problem Improving exploration in complex games.
method Episodic memory-based intrinsic reward, self-supervised inverse dynamics, UVFA framework.
result Doubles performance in hard exploration games, achieves non-zero rewards in Pitfall!.
Study on the topology of ordered disc configurations, revealing nontrivial homotopy classes.
problem Topology of ordered disc configurations and their homotopy types.
method Analysis of ordered configuration spaces of hard discs, focusing on homotopy types and nontrivial classes.
result Exhibit nontrivial classes in π_{n-3} for all n, and their persistence in deformed ambient discs.
GE finds failures in autonomous systems without domain heuristics.
problem Finding failures in autonomous systems without domain-specific heuristics.
method Adaptive stress testing using go-explore (GE) algorithm.
result GE finds failures in scenarios other RL techniques cannot solve.
Study shows challenges in reinforcement learning math problems, proposing enhancements and a hardness measure.
problem Challenges in reinforcement learning finding rare high-reward instances.
method Combining combinatorial group theory, algorithmic enhancements, and topological hardness measure.
result Resolved mathematical questions and proposed enhancements for reinforcement learning.
In this thesis I explore challenging discrete energy minimization problems that arise mainly in the context of computer vision tasks. This work motivates the use of such "hard-to-optimize" non-submodular functionals, and proposes methods and algorithms to cope with the NP-hardness of their optimization. Consequently, t…
We study the Combinatorial Pure Exploration problem with Continuous and Separable reward functions (CPE-CS) in the stochastic multi-armed bandit setting. In a CPE-CS instance, we are given several stochastic arms with unknown distributions, as well as a collection of possible decisions. Each decision has a reward accor…
This research develops an efficient reinforcement learning method for undercomplete POMDPs.
problem Learning undercomplete Partially Observable Markov Decision Processes (POMDPs) is computationally hard.
method OOM-UCB algorithm for episodic finite undercomplete POMDPs.
result Achieves optimal sample complexity of ildeO(1/ε2) for finding an ε-optimal policy. This paper proposes Self-Imitation Learning (SIL), a simple off-policy actor-critic algorithm that learns to reproduce the agent's past good decisions. This algorithm is designed to verify our hypothesis that exploiting past good experiences can indirectly drive deep exploration. Our empirical results show that SIL sig…
Paper investigates hardness of learning neural networks under manifold hypothesis.
problem Hardness of learning neural networks under the manifold hypothesis.
method Extending proofs of hardness in the SQ and cryptographic settings to the geometric setting.
result Learning is hard under input manifolds of bounded curvature but learnable with additional assumptions on manifold volume.
The paper explores Hodge decomposition and Hard Lefschetz Condition on almost Kähler manifolds.
problem Analyzing harmonic forms and Hodge decomposition on almost Kähler manifolds.
method Using Hodge decomposition and the Hard Lefschetz Condition to study almost Kähler manifolds.
result The spaces of harmonic forms have the Hodge decomposition and the Hard Lefschetz Condition is satisfied.
Memory networks are neural networks with an explicit memory component that can be both read and written to by the network. The memory is often addressed in a soft way using a softmax function, making end-to-end training with backpropagation possible. However, this is not computationally scalable for applications which …
Meta-learning strategy improves few-shot classification performance.
problem Few-shot classification with deep neural networks struggles when labeled samples are limited.
method Proposes an easy-to-hard expert meta-training strategy to arrange training tasks based on task hardness.
result Meta-learners achieve better results with the proposed expert training strategy.
New research shows deep learning struggles with hard problems due to biased data generation.
problem Deep learning's limitations in solving computationally hard problems.
method Proved that polynomial-time sample generators for NP-hard problems sample from easier sub-problems.
result Machine learning models trained on biased datasets overestimate their accuracy for hard problems.
This research explores various sampling methods and probability distributions for hard alignment in sequence-to-sequence TTS synthesis.
problem Improving alignment accuracy in sequence-to-sequence text-to-speech synthesis.
method Investigated various sampling methods (greedy, beam, random) and probability distributions (Bernoulli, Concrete) for hard alignment.
result Deterministic search is more preferable than stochastic search for natural alignment transition.
Deep reinforcement learning methods traditionally struggle with tasks where environment rewards are particularly sparse. One successful method of guiding exploration in these domains is to imitate trajectories provided by a human demonstrator. However, these demonstrations are typically collected under artificial condi…
TREK uses distillation to help students solve hard problems.
problem Stalled progress on hard prompts when current policy lacks useful reasoning trajectories.
method TREK combines distillation and reinforcement learning to expand student support.
result TREK significantly improves student performance on mathematical reasoning and agentic tasks.
Proposes EVE for efficient exploration in reinforcement learning.
problem Efficient exploration in reinforcement learning.
method EVE: a recipe for posterior over parameters, facilitating efficient exploration.
result Competitive performance on benchmarks, efficient exploration confirmed.
We study the exploration problem in episodic MDPs with rich observations generated from a small number of latent states. Under certain identifiability assumptions, we demonstrate how to estimate a mapping from the observations to latent states inductively through a sequence of regression and clustering steps -- where p…
Our work is a simple extension of the paper "Exploration by Random Network Distillation". More in detail, we show how to efficiently combine Intrinsic Rewards with Experience Replay in order to achieve more efficient and robust exploration (with respect to PPO/RND) and consequently better results in terms of agent perf…
Curriculum learning speeds up agent learning in Minecraft, a complex visual domain.
problem Training agents to learn multiple tasks in a complex, visual domain.
method Learning-progress based curriculum and dynamic exploration bonuses.
result Curriculum learning improves agent performance in a complex reinforcement learning problem.
We consider an agent's uncertainty about its environment and the problem of generalizing this uncertainty across observations. Specifically, we focus on the problem of exploration in non-tabular reinforcement learning. Drawing inspiration from the intrinsic motivation literature, we use density models to measure uncert…
Estimates task difficulty and transferability without trained models.
problem Estimating the difficulty and transferability of supervised classification tasks.
method Information theoretic approach using conditional entropy between training labels.
result Hardness and transferability estimates correlate with empirical values.
The paper explores when linear system identification is hard or easy, especially for under-actuated systems.
problem Statistical hardness of learning linear systems, especially under-actuated or under-excited systems.
method Using tools from minimax theory and recent statistical tools for finite sample analysis of system identification.
result The controllability index of linear systems affects the sample complexity of identification, making some systems hard to learn.
We introduce a learning framework called learning using privileged information (LUPI) to the computer vision field. We focus on the prototypical computer vision problem of teaching computers to recognize objects in images. We want the computers to be able to learn faster at the expense of providing extra information du…
Efficiently clusters noisy data with minimal queries.
problem Clustering elements with noisy oracle feedback.
method Combination of sampling strategy and correlation clustering algorithm.
result First polynomial-time algorithms for NP-hard optimization problem.
Hyper addresses the hyperparameter tuning challenge in RL, improving exploration efficiency and robustness.
problem Hyperparameter tuning is a significant challenge in RL, especially for curiosity-based exploration methods.
method Hyper robustly explores by effectively regularizing exploration visits and decoupling exploitation.
result Hyper is provably efficient and robust in various RL environments.
This paper explores the computational hardness of generating latent vectors for generative models.
problem Computational hardness of generating latent vectors for generative models.
method Established lower bounds for exact and approximate model inversion under strong exponential time hypothesis (SETH) and exponential time hypothesis (ETH).
result Lower bounds for computational complexity of exact and approximate model inversion.
We present an adversarial active exploration for inverse dynamics model learning, a simple yet effective learning scheme that incentivizes exploration in an environment without any human intervention. Our framework consists of a deep reinforcement learning (DRL) agent and an inverse dynamics model contesting with each …
In graph-based active learning, algorithms based on expected error minimization (EEM) have been popular and yield good empirical performance. The exact computation of EEM optimally balances exploration and exploitation. In practice, however, EEM-based algorithms employ various approximations due to the computational ha…
Novelty search in low-dimensional space improves sample efficiency in exploration tasks.
problem Efficient exploration in complex environments with sparse rewards.
method Combines model-based and model-free objectives to learn a low-dimensional representation. Uses intrinsic novelty rewards based on nearest neighbor distances in this space.
result Our approach achieves more sample-efficient exploration compared to strong baselines on various tasks.
New policy tackles evolving externalities in contextual bandits.
problem Difficulty in recovering from wrong decisions over time.
method Rejection-based policy to achieve low regret.
result Low regret achieved regardless of reward matrix structure.
ECO-DQN uses RL to improve combinatorial optimization solutions by continuous exploration.
problem Finding optimal subsets or orderings in NP-hard combinatorial optimization problems.
method ECO-DQN learns to explore continuously at test time, improving solutions incrementally.
result ECO-DQN achieves state-of-the-art performance on the Maximum Cut problem.
The paper explores when to prioritize easy or hard samples in learning tasks.
problem Determining the optimal order of learning easy or hard samples.
method Theoretical analyses and experiments were conducted to propose and validate four priority modes.
result Four priority modes (easy-first, hard-first, medium-first, two-ends-first) can be flexibly applied.
Regularization-induced exploration improves contextual bandit performance.
problem Complex reward models in real-world contextual bandits are hard to explore effectively.
method Regularization-induced exploration using stochasticity in cross-validation.
result Regularization-induced exploration leads to reliable exploration in large-scale business environments.
The application of stochastic variance reduction to optimization has shown remarkable recent theoretical and practical success. The applicability of these techniques to the hard non-convex optimization problems encountered during training of modern deep neural networks is an open problem. We show that naive application…
New framework uses unsupervised learning for efficient exploration in RL.
problem Efficient exploration in reinforcement learning with rich observations.
method Combines unsupervised learning and no-regret RL algorithms.
result Proves sample complexity for finding near-optimal policies.
The paper explores conditions for topological rigidity in quotients of the Davis complex.
problem Understanding when quotients of the Davis complex are topologically rigid.
method Analyzing quotients of the Davis complex of right-angled Coxeter groups and conditions on defining graphs.
result Introduction of infinitely many infinite topologically rigid subclasses.
In distributional reinforcement learning (RL), the estimated distribution of value function models both the parametric and intrinsic uncertainties. We propose a novel and efficient exploration method for deep RL that has two components. The first is a decaying schedule to suppress the intrinsic uncertainty. The second …
We present a learning-based approach to computing solutions for certain NP-hard problems. Our approach combines deep learning techniques with useful algorithmic elements from classic heuristics. The central component is a graph convolutional network that is trained to estimate the likelihood, for each vertex in a graph…
Surveying a new method to predict computational hardness in hypothesis testing.
problem Understanding statistical-versus-computational tradeoffs in high-dimensional inference problems.
method The low-degree method, which predicts computational hardness using the second moment of the low-degree likelihood ratio.
result Sharp low-degree lower bounds against subexponential-time algorithms for tensor PCA.
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.