Hengyuan Hu

Hengyuan Hu, Suvir Mirchandani, Dorsa Sadigh · stanford

Despite the considerable potential of reinforcement learning (RL), robotic control tasks predominantly rely on imitation learning (IL) due to its better sample efficiency. However, it is costly to collect comprehensive expert demonstrations that enable IL to generalize to all possible scenarios, and any distribution shift would require recollecting data for finetuning. Therefore, RL is appealing if it can build upon IL as an efficient autonomous self-improvement procedure. We propose imitation bootstrapped reinforcement learning (IBRL), a novel framework for sample-efficient RL with demonstrations that first trains an IL policy on the provided demonstrations and then uses it to propose alternative actions for both online exploration and bootstrapping target values. Compared to prior works that oversample the demonstrations or regularize RL with an additional imitation loss, IBRL is able to utilize high quality actions from IL policies since the beginning of training, which greatly accelerates exploration and training efficiency. We evaluate IBRL on 6 simulation and 3 real-world tasks spanning various difficulty levels. IBRL significantly outperforms prior methods and the improvement is particularly more prominent in harder tasks.

Minae Kwon, Hengyuan Hu, Vivek Myers et al. · stanford

Consider a robot tasked with tidying a desk with a meticulously constructed Lego sports car. A human may recognize that it is not appropriate to disassemble the sports car and put it away as part of the "tidying." How can a robot reach that conclusion? Although large language models (LLMs) have recently been used to enable commonsense reasoning, grounding this reasoning in the real world has been challenging. To reason in the real world, robots must go beyond passively querying LLMs and actively gather information from the environment that is required to make the right decision. For instance, after detecting that there is an occluded car, the robot may need to actively perceive the car to know whether it is an advanced model car made out of Legos or a toy car built by a toddler. We propose an approach that leverages an LLM and vision language model (VLM) to help a robot actively perceive its environment to perform grounded commonsense reasoning. To evaluate our framework at scale, we release the MessySurfaces dataset which contains images of 70 real-world surfaces that need to be cleaned. We additionally illustrate our approach with a robot on 2 carefully designed surfaces. We find an average 12.9% improvement on the MessySurfaces benchmark and an average 15% improvement on the robot experiments over baselines that do not use active perception. The dataset, code, and videos of our approach can be found at https://minaek.github.io/grounded_commonsense_reasoning.

Hengyuan Hu, Dorsa Sadigh

One of the fundamental quests of AI is to produce agents that coordinate well with humans. This problem is challenging, especially in domains that lack high quality human behavioral data, because multi-agent reinforcement learning (RL) often converges to different equilibria from the ones that humans prefer. We propose a novel framework, instructRL, that enables humans to specify what kind of strategies they expect from their AI partners through natural language instructions. We use pretrained large language models to generate a prior policy conditioned on the human instruction and use the prior to regularize the RL objective. This leads to the RL agent converging to equilibria that are aligned with human preferences. We show that instructRL converges to human-like policies that satisfy the given instructions in a proof-of-concept environment as well as the challenging Hanabi benchmark. Finally, we show that knowing the language instruction significantly boosts human-AI coordination performance in human evaluations in Hanabi.

Hengyuan Hu, Samuel Sokota, David Wu et al. · meta-ai, oxford

Fully cooperative, partially observable multi-agent problems are ubiquitous in the real world. In this paper, we focus on a specific subclass of coordination problems in which humans are able to discover self-explaining deviations (SEDs). SEDs are actions that deviate from the common understanding of what reasonable behavior would be in normal circumstances. They are taken with the intention of causing another agent or other agents to realize, using theory of mind, that the circumstance must be abnormal. We first motivate SED with a real world example and formalize its definition. Next, we introduce a novel algorithm, improvement maximizing self-explaining deviations (IMPROVISED), to perform SEDs. Lastly, we evaluate IMPROVISED both in an illustrative toy setting and the popular benchmark setting Hanabi, where it is the first method to produce so called finesse plays, which are regarded as one of the more iconic examples of human theory of mind.

Samuel Sokota, Gabriele Farina, David J. Wu et al.

The process of revising (or constructing) a policy at execution time -- known as decision-time planning -- has been key to achieving superhuman performance in perfect-information games like chess and Go. A recent line of work has extended decision-time planning to imperfect-information games, leading to superhuman performance in poker. However, these methods involve solving subgames whose sizes grow quickly in the amount of non-public information, making them unhelpful when the amount of non-public information is large. Motivated by this issue, we introduce an alternative framework for decision-time planning that is not based on solving subgames, but rather on update equivalence. In this update-equivalence framework, decision-time planning algorithms replicate the updates of last-iterate algorithms, which need not rely on public information. This facilitates scalability to games with large amounts of non-public information. Using this framework, we derive a provably sound search algorithm for fully cooperative games based on mirror descent and a search algorithm for adversarial games based on magnetic mirror descent. We validate the performance of these algorithms in cooperative and adversarial domains, notably in Hanabi, the standard benchmark for search in fully cooperative imperfect-information games. Here, our mirror descent approach exceeds or matches the performance of public information-based search while using two orders of magnitude less search time. This is the first instance of a non-public-information-based algorithm outperforming public-information-based approaches in a domain they have historically dominated.

Ifdita Hasan Orney, Jubayer Ibn Hamid, Shreya S Ramanujam et al.

Exploration is a cornerstone of learning from experience: it enables agents to find solutions to complex problems, generalize to novel ones, and scale performance with test-time compute. In this paper, we present a framework for post-training language models (LMs) that explicitly encourages optimistic exploration and promotes a synergy between exploration and exploitation. The central idea is to train the LM to generate sets of responses that are collectively accurate under the reward function and exploratory in their reasoning strategies. We first develop a general recipe for optimizing LMs with set reinforcement learning (set RL) under arbitrary objective functions, showing how standard RL algorithms can be adapted to this setting through a modification to the advantage computation. We then propose Polychromic Exploratory Policy Optimization (Poly-EPO), which instantiates this framework with an objective that explicitly synergizes exploration and exploitation. Across a range of reasoning benchmarks, we show that Poly-EPO improves generalization, as evidenced by higher pass@$k$ coverage, preserves greater diversity in model generations, and effectively scales with test-time compute.

Brandon Cui, Hengyuan Hu, Luis Pineda et al.

The standard problem setting in cooperative multi-agent settings is self-play (SP), where the goal is to train a team of agents that works well together. However, optimal SP policies commonly contain arbitrary conventions ("handshakes") and are not compatible with other, independently trained agents or humans. This latter desiderata was recently formalized by Hu et al. 2020 as the zero-shot coordination (ZSC) setting and partially addressed with their Other-Play (OP) algorithm, which showed improved ZSC and human-AI performance in the card game Hanabi. OP assumes access to the symmetries of the environment and prevents agents from breaking these in a mutually incompatible way during training. However, as the authors point out, discovering symmetries for a given environment is a computationally hard problem. Instead, we show that through a simple adaption of k-level reasoning (KLR) Costa Gomes et al. 2006, synchronously training all levels, we can obtain competitive ZSC and ad-hoc teamplay performance in Hanabi, including when paired with a human-like proxy bot. We also introduce a new method, synchronous-k-level reasoning with a best response (SyKLRBR), which further improves performance on our synchronous KLR by co-training a best response.

Hengyuan Hu, David J Wu, Adam Lerer et al.

We consider the problem of making AI agents that collaborate well with humans in partially observable fully cooperative environments given datasets of human behavior. Inspired by piKL, a human-data-regularized search method that improves upon a behavioral cloning policy without diverging far away from it, we develop a three-step algorithm that achieve strong performance in coordinating with real humans in the Hanabi benchmark. We first use a regularized search algorithm and behavioral cloning to produce a better human model that captures diverse skill levels. Then, we integrate the policy regularization idea into reinforcement learning to train a human-like best response to the human model. Finally, we apply regularized search on top of the best response policy at test time to handle out-of-distribution challenges when playing with humans. We evaluate our method in two large scale experiments with humans. First, we show that our method outperforms experts when playing with a group of diverse human players in ad-hoc teams. Second, we show that our method beats a vanilla best response to behavioral cloning baseline by having experts play repeatedly with the two agents.

Samuel Sokota, Eugene Vinitsky, Hengyuan Hu et al.

Few classical games have been regarded as such significant benchmarks of artificial intelligence as to have justified training costs in the millions of dollars. Among these, Stratego -- a board wargame exemplifying the challenge of strategic decision making under massive amounts of hidden information -- stands apart as a case where such efforts failed to produce performance at the level of top humans. This work establishes a step change in both performance and cost for Stratego, showing that it is now possible not only to reach the level of top humans, but to achieve vastly superhuman level -- and that doing so requires not an industrial budget, but merely a few thousand dollars. We achieved this result by developing general approaches for self-play reinforcement learning and test-time search under imperfect information.

Hengyuan Hu, Tingchen Fu, Minqi Jiang et al.

Recent years have witnessed tremendous progress in enabling LLMs to solve complex reasoning tasks such as math and coding. As we start to apply LLMs to harder tasks that they may not be able to solve in one shot, it is worth paying attention to their ability to construct intermediate stepping stones that prepare them to better solve the tasks. Examples of stepping stones include simplifications, alternative framings, or subproblems. We study properties and benefits of stepping stones in the context of modern reasoning LLMs via ARQ (\textbf{A}king the \textbf{R}ight \textbf{Q}uestions), our simple framework which introduces a question generator to the default reasoning pipeline. We first show that good stepping stone questions exist and are transferrable, meaning that good questions can be generated, and they substantially help LLMs of various capabilities in solving the target tasks. We next frame stepping stone generation as a post-training task and show that we can fine-tune LLMs to generate more useful stepping stones by SFT and RL on synthetic data.

Priya Sundaresan, Hengyuan Hu, Quan Vuong et al.

While imitation learning (IL) offers a promising framework for teaching robots various behaviors, learning complex tasks remains challenging. Existing IL policies struggle to generalize effectively across visual and spatial variations even for simple tasks. In this work, we introduce SPHINX: Salient Point-based Hybrid ImitatioN and eXecution, a flexible IL policy that leverages multimodal observations (point clouds and wrist images), along with a hybrid action space of low-frequency, sparse waypoints and high-frequency, dense end effector movements. Given 3D point cloud observations, SPHINX learns to infer task-relevant points within a point cloud, or salient points, which support spatial generalization by focusing on semantically meaningful features. These salient points serve as anchor points to predict waypoints for long-range movement, such as reaching target poses in free-space. Once near a salient point, SPHINX learns to switch to predicting dense end-effector movements given close-up wrist images for precise phases of a task. By exploiting the strengths of different input modalities and action representations for different manipulation phases, SPHINX tackles complex tasks in a sample-efficient, generalizable manner. Our method achieves 86.7% success across 4 real-world and 2 simulated tasks, outperforming the next best state-of-the-art IL baseline by 41.1% on average across 440 real world trials. SPHINX additionally generalizes to novel viewpoints, visual distractors, spatial arrangements, and execution speeds with a 1.7x speedup over the most competitive baseline. Our website (http://sphinx-manip.github.io) provides open-sourced code for data collection, training, and evaluation, along with supplementary videos.

Hengyuan Hu, Jakob N Foerster

In recent years we have seen fast progress on a number of benchmark problems in AI, with modern methods achieving near or super human performance in Go, Poker and Dota. One common aspect of all of these challenges is that they are by design adversarial or, technically speaking, zero-sum. In contrast to these settings, success in the real world commonly requires humans to collaborate and communicate with others, in settings that are, at least partially, cooperative. In the last year, the card game Hanabi has been established as a new benchmark environment for AI to fill this gap. In particular, Hanabi is interesting to humans since it is entirely focused on theory of mind, i.e., the ability to effectively reason over the intentions, beliefs and point of view of other agents when observing their actions. Learning to be informative when observed by others is an interesting challenge for Reinforcement Learning (RL): Fundamentally, RL requires agents to explore in order to discover good policies. However, when done naively, this randomness will inherently make their actions less informative to others during training. We present a new deep multi-agent RL method, the Simplified Action Decoder (SAD), which resolves this contradiction exploiting the centralized training phase. During training SAD allows other agents to not only observe the (exploratory) action chosen, but agents instead also observe the greedy action of their team mates. By combining this simple intuition with best practices for multi-agent learning, SAD establishes a new SOTA for learning methods for 2-5 players on the self-play part of the Hanabi challenge. Our ablations show the contributions of SAD compared with the best practice components. All of our code and trained agents are available at https://github.com/facebookresearch/Hanabi_SAD.

Hengyuan Hu, Aniket Das, Dorsa Sadigh et al.

Denoising Diffusion Probabilistic Models (DDPMs) have emerged as powerful tools for generative modeling. However, their sequential computation requirements lead to significant inference-time bottlenecks. In this work, we utilize the connection between DDPMs and Stochastic Localization to prove that, under an appropriate reparametrization, the increments of DDPM satisfy an exchangeability property. This general insight enables near-black-box adaptation of various performance optimization techniques from autoregressive models to the diffusion setting. To demonstrate this, we introduce \emph{Autospeculative Decoding} (ASD), an extension of the widely used speculative decoding algorithm to DDPMs that does not require any auxiliary draft models. Our theoretical analysis shows that ASD achieves a $\tilde{O} (K^{\frac{1}{3}})$ parallel runtime speedup over the $K$ step sequential DDPM. We also demonstrate that a practical implementation of autospeculative decoding accelerates DDPM inference significantly in various domains.

Athul Paul Jacob, David J. Wu, Gabriele Farina et al.

We consider the task of building strong but human-like policies in multi-agent decision-making problems, given examples of human behavior. Imitation learning is effective at predicting human actions but may not match the strength of expert humans, while self-play learning and search techniques (e.g. AlphaZero) lead to strong performance but may produce policies that are difficult for humans to understand and coordinate with. We show in chess and Go that regularizing search based on the KL divergence from an imitation-learned policy results in higher human prediction accuracy and stronger performance than imitation learning alone. We then introduce a novel regret minimization algorithm that is regularized based on the KL divergence from an imitation-learned policy, and show that using this algorithm for search in no-press Diplomacy yields a policy that matches the human prediction accuracy of imitation learning while being substantially stronger.

Arnaud Fickinger, Hengyuan Hu, Brandon Amos et al.

Lookahead search has been a critical component of recent AI successes, such as in the games of chess, go, and poker. However, the search methods used in these games, and in many other settings, are tabular. Tabular search methods do not scale well with the size of the search space, and this problem is exacerbated by stochasticity and partial observability. In this work we replace tabular search with online model-based fine-tuning of a policy neural network via reinforcement learning, and show that this approach outperforms state-of-the-art search algorithms in benchmark settings. In particular, we use our search algorithm to achieve a new state-of-the-art result in self-play Hanabi, and show the generality of our algorithm by also showing that it outperforms tabular search in the Atari game Ms. Pacman.

Hengyuan Hu, Adam Lerer, Noam Brown et al.

Search is an important tool for computing effective policies in single- and multi-agent environments, and has been crucial for achieving superhuman performance in several benchmark fully and partially observable games. However, one major limitation of prior search approaches for partially observable environments is that the computational cost scales poorly with the amount of hidden information. In this paper we present \emph{Learned Belief Search} (LBS), a computationally efficient search procedure for partially observable environments. Rather than maintaining an exact belief distribution, LBS uses an approximate auto-regressive counterfactual belief that is learned as a supervised task. In multi-agent settings, LBS uses a novel public-private model architecture for underlying policies in order to efficiently evaluate these policies during rollouts. In the benchmark domain of Hanabi, LBS can obtain 55% ~ 91% of the benefit of exact search while reducing compute requirements by $35.8 \times$ ~ $4.6 \times$, allowing it to scale to larger settings that were inaccessible to previous search methods.

Hengyuan Hu, Adam Lerer, Brandon Cui et al.

The standard problem setting in Dec-POMDPs is self-play, where the goal is to find a set of policies that play optimally together. Policies learned through self-play may adopt arbitrary conventions and implicitly rely on multi-step reasoning based on fragile assumptions about other agents' actions and thus fail when paired with humans or independently trained agents at test time. To address this, we present off-belief learning (OBL). At each timestep OBL agents follow a policy $π_1$ that is optimized assuming past actions were taken by a given, fixed policy ($π_0$), but assuming that future actions will be taken by $π_1$. When $π_0$ is uniform random, OBL converges to an optimal policy that does not rely on inferences based on other agents' behavior (an optimal grounded policy). OBL can be iterated in a hierarchy, where the optimal policy from one level becomes the input to the next, thereby introducing multi-level cognitive reasoning in a controlled manner. Unlike existing approaches, which may converge to any equilibrium policy, OBL converges to a unique policy, making it suitable for zero-shot coordination (ZSC). OBL can be scaled to high-dimensional settings with a fictitious transition mechanism and shows strong performance in both a toy-setting and the benchmark human-AI & ZSC problem Hanabi.

Jack Parker-Holder, Luke Metz, Cinjon Resnick et al.

Over the last decade, a single algorithm has changed many facets of our lives - Stochastic Gradient Descent (SGD). In the era of ever decreasing loss functions, SGD and its various offspring have become the go-to optimization tool in machine learning and are a key component of the success of deep neural networks (DNNs). While SGD is guaranteed to converge to a local optimum (under loose assumptions), in some cases it may matter which local optimum is found, and this is often context-dependent. Examples frequently arise in machine learning, from shape-versus-texture-features to ensemble methods and zero-shot coordination. In these settings, there are desired solutions which SGD on 'standard' loss functions will not find, since it instead converges to the 'easy' solutions. In this paper, we present a different approach. Rather than following the gradient, which corresponds to a locally greedy direction, we instead follow the eigenvectors of the Hessian, which we call "ridges". By iteratively following and branching amongst the ridges, we effectively span the loss surface to find qualitatively different solutions. We show both theoretically and experimentally that our method, called Ridge Rider (RR), offers a promising direction for a variety of challenging problems.

Hengyuan Hu, Adam Lerer, Alex Peysakhovich et al.

We consider the problem of zero-shot coordination - constructing AI agents that can coordinate with novel partners they have not seen before (e.g. humans). Standard Multi-Agent Reinforcement Learning (MARL) methods typically focus on the self-play (SP) setting where agents construct strategies by playing the game with themselves repeatedly. Unfortunately, applying SP naively to the zero-shot coordination problem can produce agents that establish highly specialized conventions that do not carry over to novel partners they have not been trained with. We introduce a novel learning algorithm called other-play (OP), that enhances self-play by looking for more robust strategies, exploiting the presence of known symmetries in the underlying problem. We characterize OP theoretically as well as experimentally. We study the cooperative card game Hanabi and show that OP agents achieve higher scores when paired with independently trained agents. In preliminary results we also show that our OP agents obtains higher average scores when paired with human players, compared to state-of-the-art SP agents.

Tristan Cazenave, Yen-Chi Chen, Guan-Wei Chen et al.

Since DeepMind's AlphaZero, Zero learning quickly became the state-of-the-art method for many board games. It can be improved using a fully convolutional structure (no fully connected layer). Using such an architecture plus global pooling, we can create bots independent of the board size. The training can be made more robust by keeping track of the best checkpoints during the training and by training against them. Using these features, we release Polygames, our framework for Zero learning, with its library of games and its checkpoints. We won against strong humans at the game of Hex in 19x19, which was often said to be untractable for zero learning; and in Havannah. We also won several first places at the TAAI competitions.

Adam Lerer, Hengyuan Hu, Jakob Foerster et al.

Recent superhuman results in games have largely been achieved in a variety of zero-sum settings, such as Go and Poker, in which agents need to compete against others. However, just like humans, real-world AI systems have to coordinate and communicate with other agents in cooperative partially observable environments as well. These settings commonly require participants to both interpret the actions of others and to act in a way that is informative when being interpreted. Those abilities are typically summarized as theory f mind and are seen as crucial for social interactions. In this paper we propose two different search techniques that can be applied to improve an arbitrary agreed-upon policy in a cooperative partially observable game. The first one, single-agent search, effectively converts the problem into a single agent setting by making all but one of the agents play according to the agreed-upon policy. In contrast, in multi-agent search all agents carry out the same common-knowledge search procedure whenever doing so is computationally feasible, and fall back to playing according to the agreed-upon policy otherwise. We prove that these search procedures are theoretically guaranteed to at least maintain the original performance of the agreed-upon policy (up to a bounded approximation error). In the benchmark challenge problem of Hanabi, our search technique greatly improves the performance of every agent we tested and when applied to a policy trained using RL achieves a new state-of-the-art score of 24.61 / 25 in the game, compared to a previous-best of 24.08 / 25.

Hengyuan Hu, Denis Yarats, Qucheng Gong et al.

We explore using latent natural language instructions as an expressive and compositional representation of complex actions for hierarchical decision making. Rather than directly selecting micro-actions, our agent first generates a latent plan in natural language, which is then executed by a separate model. We introduce a challenging real-time strategy game environment in which the actions of a large number of units must be coordinated across long time scales. We gather a dataset of 76 thousand pairs of instructions and executions from human play, and train instructor and executor models. Experiments show that models using natural language as a latent variable significantly outperform models that directly imitate human actions. The compositional structure of language proves crucial to its effectiveness for action representation. We also release our code, models and data.

Hengyuan Hu, Lisheng Gao, Quanbin Ma

Building a good generative model for image has long been an important topic in computer vision and machine learning. Restricted Boltzmann machine (RBM) is one of such models that is simple but powerful. However, its restricted form also has placed heavy constraints on the models representation power and scalability. Many extensions have been invented based on RBM in order to produce deeper architectures with greater power. The most famous ones among them are deep belief network, which stacks multiple layer-wise pretrained RBMs to form a hybrid model, and deep Boltzmann machine, which allows connections between hidden units to form a multi-layer structure. In this paper, we present a new method to compose RBMs to form a multi-layer network style architecture and a training method that trains all layers jointly. We call the resulted structure deep restricted Boltzmann network. We further explore the combination of convolutional RBM with the normal fully connected RBM, which is made trivial under our composition framework. Experiments show that our model can generate descent images and outperform the normal RBM significantly in terms of image quality and feature quality, without losing much efficiency for training.

Hengyuan Hu, Rui Peng, Yu-Wing Tai et al.

State-of-the-art neural networks are getting deeper and wider. While their performance increases with the increasing number of layers and neurons, it is crucial to design an efficient deep architecture in order to reduce computational and memory costs. Designing an efficient neural network, however, is labor intensive requiring many experiments, and fine-tunings. In this paper, we introduce network trimming which iteratively optimizes the network by pruning unimportant neurons based on analysis of their outputs on a large dataset. Our algorithm is inspired by an observation that the outputs of a significant portion of neurons in a large network are mostly zero, regardless of what inputs the network received. These zero activation neurons are redundant, and can be removed without affecting the overall accuracy of the network. After pruning the zero activation neurons, we retrain the network using the weights before pruning as initialization. We alternate the pruning and retraining to further reduce zero activations in a network. Our experiments on the LeNet and VGG-16 show that we can achieve high compression ratio of parameters without losing or even achieving higher accuracy than the original network.