A fundamental challenge in multi-agent reinforcement learning (MARL) is to learn the joint policy in an extremely large search space, which grows exponentially with the number of agents. Moreover, fully decentralized policy factorization significantly restricts the search space, which may lead to sub-optimal policies. In contrast, the auto-regressive joint policy can represent a much richer class of joint policies by factorizing the joint policy into the product of a series of conditional individual policies. While such factorization introduces the action dependency among agents explicitly in sequential execution, it does not take full advantage of the dependency during learning. In particular, the subsequent agents do not give the preceding agents feedback about their decisions. In this paper, we propose a new framework Back-Propagation Through Agents (BPTA) that directly accounts for both agents' own policy updates and the learning of their dependent counterparts. This is achieved by propagating the feedback through action chains. With the proposed framework, our Bidirectional Proximal Policy Optimisation (BPPO) outperforms the state-of-the-art methods. Extensive experiments on matrix games, StarCraftII v2, Multi-agent MuJoCo, and Google Research Football demonstrate the effectiveness of the proposed method.
论文检索
输入标题、作者或关键词，从 510 篇学术成果中精准定位
Proximal Policy Optimization algorithm employing a clipped surrogate objective (PPO-Clip) is a prominent exemplar of the policy optimization methods. However, despite its remarkable empirical success, PPO-Clip lacks theoretical substantiation to date. In this paper, we contribute to the field by establishing the first global convergence results of a PPO-Clip variant in both tabular and neural function approximation settings. Our findings highlight the O(1/√T ) min-iterate convergence rate specifically in the context of neural function approximation. We tackle the inherent challenges in analyzing PPO-Clip through three central concepts: (i) We introduce a generalized version of the PPO-Clip objective, illuminated by its connection with the hinge loss. (ii) Employing entropic mirror descent, we establish asymptotic convergence for tabular PPO-Clip with direct policy parameterization. (iii) Inspired by the tabular analysis, we streamline convergence analysis by introducing a two-step policy improvement approach. This decouples policy search from complex neural policy parameterization using a regression-based update scheme. Furthermore, we gain deeper insights into the efficacy of PPO-Clip by interpreting these generalized objectives. Our theoretical findings also mark the first characterization of the influence of the clipping mechanism on PPO-Clip convergence. Importantly, the clipping range affects only the pre-constant of the convergence rate.
AlphaZero and MuZero have achieved state-of-the-art (SOTA) performance in a wide range of domains, including board games and robotics, with discrete and continuous action spaces. However, to obtain an improved policy, they often require an excessively large number of simulations, especially for domains with large action spaces. As the simulation budget decreases, their performance drops significantly. In addition, many important real-world applications have combinatorial (or exponential) action spaces, making it infeasible to search directly over all possible actions. In this paper, we extend AlphaZero and MuZero to learn and plan in more complex multiagent (MA) Markov decision processes, where the action spaces increase exponentially with the number of agents. Our new algorithms, MA Gumbel AlphaZero and MA Gumbel MuZero, respectively without and with model learning, achieve superior performance on cooperative multiagent control problems, while reducing the number of environmental interactions by up to an order of magnitude compared to model-free approaches. In particular, we significantly improve prior performance when planning with much fewer simulation budgets. The code and appendix are available at https://github.com/tjuHaoXiaotian/MA-MuZero.
Meta-reinforcement learning (meta-RL) aims to quickly solve new RL tasks by leveraging knowledge from prior tasks. Previous studies often assume a single-mode homogeneous task distribution, ignoring possible structured heterogeneity among tasks. Such an oversight can hamper effective exploration and adaptation, especially with limited samples. In this work, we harness the structured heterogeneity among tasks via clustering to improve meta-RL, which facilitates knowledge sharing at the cluster level. To facilitate exploration, we also develop a dedicated cluster-level exploratory policy to discover task clusters via divide-and-conquer. The knowledge from the discovered clusters helps to narrow the search space of task-specific policy learning, leading to more sample-efficient policy adaptation. We evaluate the proposed method on environments with parametric clusters (e.g., rewards and state dynamics in the MuJoCo suite) and non-parametric clusters (e.g., control skills in the Meta-World suite). The results demonstrate strong advantages of our solution against a set of representative meta-RL methods.
Offline optimization is an emerging problem in many experimental engineering domains including protein, drug or aircraft design, where online experimentation to collect evaluation data is too expensive or dangerous. To avoid that, one has to optimize an unknown function given only its offline evaluation at a fixed set of inputs. A naive solution to this problem is to learn a surrogate model of the unknown function and optimize this surrogate instead. However, such a naive optimizer is prone to erroneous overestimation of the surrogate (possibly due to over-fitting on a biased sample of function evaluation) on inputs outside the offline dataset. Prior approaches addressing this challenge have primarily focused on learning robust surrogate models. However, their search strategies are derived from the surrogate model rather than the actual offline data. To fill this important gap, we introduce a new learning-to-search perspective for offline optimization by reformulating it as an offline reinforcement learning problem. Our proposed policy-guided gradient search approach explicitly learns the best policy for a given surrogate model created from the offline data. Our empirical results on multiple benchmarks demonstrate that the learned optimization policy can be combined with existing offline surrogates to significantly improve the optimization performance.
Program synthesis is the task of automatically generating code based on a specification. In Syntax-Guided Synthesis (SyGuS) this specification is a combination of a syntactic template and a logical formula, and the result is guaranteed to satisfy both. We present a reinforcement-learning guided algorithm for SyGuS which uses Monte-Carlo Tree Search (MCTS) to search the space of candidate solutions. Our algorithm learns policy and value functions which, combined with the upper confidence bound for trees, allow it to balance exploration and exploitation. A common challenge in applying machine learning approaches to syntax-guided synthesis is the scarcity of training data. To address this, we present a method for automatically generating training data for SyGuS based on anti-unification of existing first-order satisfiability problems, which we use to train our MCTS policy. We implement and evaluate this setup and demonstrate that learned policy and value improve the synthesis performance over a baseline by over 26 percentage points in the training and testing sets. Our tool outperforms state-of-the-art tool cvc5 on the training set and performs comparably in terms of the total number of problems solved on the testing set (solving 23% of the benchmarks on which cvc5 fails). We make our data set publicly available, to enable further application of machine learning methods to the SyGuS problem.
Conflict-driven clause learning (CDCL) is the dominating algorithmic paradigm for SAT solving and hugely successful in practice. In its lifted version QCDCL, it is one of the main approaches for solving quantified Boolean formulas (QBF). In both SAT and QBF, proofs can be efficiently extracted from runs of (Q)CDCL solvers. While for CDCL, it is known that the proof size in the underlying proof system propositional resolution matches the CDCL runtime up to a polynomial factor, we show that in QBF there is an exponential gap between QCDCL runtime and the size of the extracted proofs in QBF resolution systems. We demonstrate that this is not just a gap between QCDCL runtime and the size of any QBF resolution proof, but even the extracted proofs are exponentially smaller for some instances. Hence searching for a small proof via QCDCL (even with non-deterministic decision policies) will provably incur an exponential overhead for some instances.
Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving
PDF ↗Diffusion models excel at modeling complex and multimodal trajectory distributions for decision-making and control. Reward-gradient guided denoising has been recently proposed to generate trajectories that maximize both a differentiable reward function and the likelihood under the data distribution captured by a diffusion model. Reward-gradient guided denoising requires a differentiable reward function fitted to both clean and noised samples limiting its applicability as a general trajectory optimizer. In this paper we propose DiffusionES a method that combines gradient-free optimization with trajectory denoising to optimize black-box non-differentiable objectives while staying in the data manifold. Diffusion-ES samples trajectories during evolutionary search from a diffusion model and scores them using a black-box reward function. It mutates high-scoring trajectories using a truncated diffusion process that applies a small number of noising and denoising steps allowing for much more efficient exploration of the solution space. We show that DiffusionES achieves state-of-the-art performance on nuPlan an established closed-loop planning benchmark for autonomous driving. Diffusion-ES outperforms existing sampling-based planners reactive deterministic or diffusion-based policies and reward-gradient guidance. Additionally we show that unlike prior guidance methods our method can optimize non-differentiable language-shaped reward functions generated by few-shot LLM prompting. When guided by a human teacher that issues instructions to follow our method can generate novel highly complex behaviors such as aggressive lane weaving which are not present in the training data. This allows us to solve the hardest nuPlan scenarios which are beyond the capabilities of existing trajectory optimization methods and driving policies.
Experience as Source for Anticipation and Planning: Experiential Policy Learning for Target-driven Recommendation Dialogues
PDF ↗Target-driven recommendation dialogues present unique challenges in dialogue management due to the necessity of anticipating user interactions for successful conversations. Current methods face significant limitations: (I) inadequate capabilities for conversation anticipation, (II) computational inefficiencies due to costly simulations, and (III) neglect of valuable past dialogue experiences. To address these limitations, we propose a new framework, Experiential Policy Learning (EPL), for enhancing such dialogues. EPL embodies the principle of Learning From Experience, facilitating anticipation with an experiential scoring function that estimates dialogue state potential using similar past interactions stored in long-term memory. To demonstrate its flexibility, we introduce Tree-structured EPL (T-EPL) as one possible training-free realization with Large Language Models (LLMs) and Monte-Carlo Tree Search (MCTS). T-EPL assesses past dialogue states with LLMs while utilizing MCTS to achieve hierarchical and multi-level reasoning. Extensive experiments on two published datasets demonstrate the superiority and efficacy of T-EPL.
Reward-based finetuning is crucial for aligning language policies with intended behaviors (*e.g.*, creativity and safety). A key challenge is to develop steerable language models that trade-off multiple (conflicting) objectives in a flexible and efficient manner. This paper presents Conditional Language Policy (CLP), a general framework for finetuning language models on multiple objectives. Building on techniques from multi-task training and parameter-efficient finetuning, CLP learn steerable models that effectively trade-off conflicting objectives at *inference time*. Notably, this does not require training or maintaining multiple models to achieve different trade-offs between the objectives. Through extensive experiments and ablations on two summarization datasets, we show that CLP learns steerable language models that outperform and Pareto-dominate the existing approaches for multi-objective
Large Language Models (LLMs) have shown great potential in the biomedical domain with the advancement of retrieval-augmented generation (RAG). However, existing retrieval-augmented approaches face challenges in addressing diverse queries and documents, particularly for medical knowledge queries, resulting in sub-optimal performance. To address these limitations, we propose a novel plug-and-play LLM-based retrieval method called Self-Rewarding Tree Search (SeRTS) based on Monte Carlo Tree Search (MCTS) and a self-rewarding paradigm. By combining the reasoning capabilities of LLMs with the effectiveness of tree search, SeRTS boosts the zero-shot performance of retrieving high-quality and informative results for RAG. We further enhance retrieval performance by fine-tuning LLMs with Proximal Policy Optimization (PPO) objectives using the trajectories collected by SeRTS as feedback. Controlled experiments using the BioASQ-QA dataset with GPT-3.5-Turbo and LLama2-7b demonstrate that our method significantly improves the performance of the BM25 retriever and surpasses the strong baseline of self-reflection in both efficiency and scalability. Moreover, SeRTS generates higher-quality feedback for PPO training than self-reflection. Our proposed method effectively adapts LLMs to document retrieval tasks, enhancing their ability to retrieve highly relevant documents for RAG in the context of medical knowledge queries. This work presents a significant step forward in leveraging LLMs for accurate and comprehensive biomedical question answering.
In conversational AI, large-scale models excel in various tasks but struggle with target-driven conversation planning. Current methods, such as chain-of-thought reasoning and tree-search policy learning techniques, either neglect plan rationality or require extensive human simulation procedures. Addressing this, we propose a novel two-stage framework, named EnPL, to improve the LLMs’ capability in planning conversations towards designated targets, including (1) distilling natural language plans from target-driven conversation corpus and (2) generating new plans with demonstration-guided in-context learning. Specifically, we first propose a filter approach to distill a high-quality plan dataset, ConvPlan (Resources of this paper can be found at https://github.com/pandazzh2020/ConvPlan). With the aid of corresponding conversational data and support from relevant knowledge bases, we validate the quality and rationality of these plans. Then, these plans are leveraged to help guide LLMs to further plan for new targets. Empirical results demonstrate that our method significantly improves the planning ability of LLMs, especially in target-driven conversations. Furthermore, EnPL is demonstrated to be quite effective in collecting target-driven conversation datasets and enhancing response generation, paving the way for constructing extensive target-driven conversational models.
Finding appropriate prompts for the specific task has become an important issue as the usage of Large Language Models (LLM) have expanded. However, the variety of input-output formats complicate finding the prompts. Reinforcement Learning (RL) is a promising for prompt tuning due to its ability to incrementally produce better results through interaction with the environment. But its inherent training instability and environmental dependency make it difficult to use in practice. In this paper, we propose StablePrompt, a prompt tuning method based on RL. We formulate prompt tuning as RL problem between agent and target LLM, and introduce Adaptive Proximal Policy Optimization (APPO), an modified version of PPO for prompt tuning. APPO introduces an anchor model and updates it adaptively based on the training trajectory. Using this anchor model for the KL divergence term in PPO keeps the search space flexible and ensures training stability. We evaluate StablePrompt on various tasks, including text classification, question answering, and text generation. StablePrompt achieves State-of-The-Art performance across diverse tasks. We demonstrates that StablePrompt performs well across various types and sizes of LLMs. Furthermore, we present TTE-StablePrompt, an extension for generating input-dependent prompts. It outperforms StablePrompt in tasks that are hard to solve with a single prompt. This shows that StablePrompt is an extensible and stable RL framework for LLM.
Graph Neural Networks
Graph Neural Networks (GNNs) have significantly advanced the field of drug discovery, enhancing the speed and efficiency of molecular identification. However, training these GNNs demands vast amounts of molecular data, which has spurred the emergence of collaborative model-sharing initiatives. These initiatives facilitate the sharing of molecular pre-trained models among organizations without exposing proprietary training data. Despite the benefits, these molecular pre-trained models may still pose privacy risks. For example, malicious adversaries could perform data extraction attack to recover private training data, thereby threatening commercial secrets and collaborative trust. This work, for the first time, explores the risks of extracting private training molecular data from molecular pre-trained models. This task is nontrivial as the molecular pre-trained models are non-generative and exhibit a diversity of model architectures, which differs significantly from language and image models. To address these issues, we introduce a molecule generation approach and propose a novel, model-independent scoring function for selecting promising molecules. To efficiently reduce the search space of potential molecules, we further introduce a Molecule Extraction Policy Network for molecule extraction. Our experiments demonstrate that even with only query access to molecular pre-trained models, there is a considerable risk of extracting training data, challenging the assumption that model sharing alone provides adequate protection against data extraction attacks. Our codes are publicly available at: \url{https://github.com/renH2/Molextract}.
Reinforcement Learning · Deep RL
Although recent advancements in large language models (LLMs) have significantly improved their performance on various tasks, they still face challenges with complex and symbolic multi-step reasoning, particularly in mathematical reasoning. To bolster the mathematical reasoning capabilities of LLMs, most existing efforts concentrate on seeking assistance from either domain experts or GPT-4 for high-quality process-supervised data, which is not only expensive but also labor-intensive. In our study, we propose an innovative framework, AlphaMath, that bypasses the need for process annotations (from humans or GPTs) by leveraging Monte Carlo Tree Search (MCTS). This framework focuses on unleashing the potential of a well-pretrained LLM to autonomously enhance its mathematical reasoning. Specifically, we integrate a value model with the LLM, automatically generating both process supervision and step-level evaluation signals in MCTS. Furthermore, we propose an efficient inference strategy—step-level beam search, where the value model is crafted to assist the policy model (i.e., LLM) in navigating more effective reasoning paths, rather than solely relying on prior probabilities. The experimental results on both in-domain and out-of-domain datasets demonstrate that even without GPT-4 or human-annotated process supervision, our AlphaMath framework achieves comparable or superior results to previous state-of-the-art methods.
Robotics
Reinforcement Learning (RL) controllers have demonstrated remarkable performance in complex robot control tasks. However, the presence of reality gap often leads to poor performance when deploying policies trained in simulation directly onto real robots. Previous sim-to-real algorithms like Domain Randomization (DR) requires domain-specific expertise and suffers from issues such as reduced control performance and high training costs. In this work, we introduce Evolutionary Adversarial Simulator Identification (EASI), a novel approach that combines Generative Adversarial Network (GAN) and Evolutionary Strategy (ES) to address sim-to-real challenges. Specifically, we consider the problem of sim-to-real as a search problem, where ES acts as a generator in adversarial competition with a neural network discriminator, aiming to find physical parameter distributions that make the state transitions between simulation and reality as similar as possible. The discriminator serves as the fitness function, guiding the evolution of the physical parameter distributions. EASI features simplicity, low cost, and high fidelity, enabling the construction of a more realistic simulator with minimal requirements for real-world data, thus aiding in transferring simulated-trained policies to the real world. We demonstrate the performance of EASI in both sim-to-sim and sim-to-real tasks, showing superior performance compared to existing sim-to-real algorithms.
Generative Models · Reasoning
Tool-augmented large language models (LLMs) leverage tools, often in the form of APIs, to improve their reasoning capabilities on complex tasks. This enables them to act as intelligent agents interacting with the real world. The recently introduced ToolLLaMA model by Qin et al. [2023] utilizes the depth-first search-based decision tree (DFSDT) mechanism for multi-step reasoning with $16000+$ real-world APIs, effectively enhancing the performance of tool-augmented LLMs compared to traditional chain reasoning mechanisms. However, their approach only employs successful paths from decision trees (also called inference trees) for supervised fine-tuning (SFT), missing out on the potential learning opportunities from failed paths. Inspired by this, we propose an inference trajectory optimization framework based on preference learning to address this limitation. We first introduce a novel method for constructing step-wise preference data from tree-like expert trajectories, which leverages the previously ignored failed explorations in the decision trees. In the subsequent training phase, we first fine-tune the LLM with successful tool-usage expert trajectories and then apply direct preference optimization (DPO) with the preference data to update the LLM's policy, resulting in our ToolPrefer-LLaMA (TP-LLaMA) model. This approach not only enhances the utilization of original expert data but also broadens the learning space of the model. Our experiments demonstrate that by obtaining insights from errors in inference trees, TP-LLaMA significantly outperforms the baselines across almost all test scenarios by a large margin and exhibits better generalization capabilities with unseen APIs. At the same time, TP-LLaMA has also demonstrated superior reasoning efficiency compared to the baselines, making it more suitable for complex tool-usage reasoning tasks.
Reinforcement Learning · Planning
Planning in real-world settings often entails addressing partial observability while aligning with users' requirements. We present a novel framework for expressing users' constraints and preferences about agent behavior in a partially observable setting using parameterized belief-state query (BSQ) policies in the setting of goal-oriented partially observable Markov decision processes (gPOMDPs). We present the first formal analysis of such constraints and prove that while the expected cost function of a parameterized BSQ policy w.r.t its parameters is not convex, it is piecewise constant and yields an implicit discrete parameter search space that is finite for finite horizons. This theoretical result leads to novel algorithms that optimize gPOMDP agent behavior with guaranteed user alignment. Analysis proves that our algorithms converge to the optimal user-aligned behavior in the limit. Empirical results show that parameterized BSQ policies provide a computationally feasible approach for user-aligned planning in partially observable settings.
Deep Learning · Everything Else
This paper uses chess, a landmark planning problem in AI, to assess transformers’ performance on a planning task where memorization is futile — even at a large scale. To this end, we release ChessBench, a large-scale benchmark dataset of 10 million chess games with legal move and value annotations (15 billion data points) provided by Stockfish 16, the state-of-the-art chess engine. We train transformers with up to 270 million parameters on ChessBench via supervised learning and perform extensive ablations to assess the impact of dataset size, model size, architecture type, and different prediction targets (state-values, action-values, and behavioral cloning). Our largest models learn to predict action-values for novel boards quite accurately, implying highly non-trivial generalization. Despite performing no explicit search, our resulting chess policy solves challenging chess puzzles and achieves a surprisingly strong Lichess blitz Elo of 2895 against humans (grandmaster level). We also compare to Leela Chess Zero and AlphaZero (trained without supervision via self-play) with and without search. We show that, although a remarkably good approximation of Stockfish’s search-based algorithm can be distilled into large-scale transformers via supervised learning, perfect distillation is still beyond reach, thus making ChessBench well-suited for future research.
How did humanity coax mathematics from the aether? We explore the Platonic view that mathematics can be discovered from its axioms---a game of conjecture and proof. We describe an agent that jointly learns to pose challenging problems for itself (conjecturing) and solve them (theorem proving). Given a mathematical domain axiomatized in dependent type theory, we first combine methods for constrained decoding and type-directed synthesis to sample valid conjectures from a language model. Our method guarantees well-formed conjectures by construction, even as we start with a randomly initialized model. We use the same model to represent a policy and value function for guiding proof search. Our agent targets generating hard but provable conjectures --- a moving target, since its own theorem proving ability also improves as it trains. We propose novel methods for hindsight relabeling on proof search trees to significantly improve the agent's sample efficiency in both tasks. Experiments on 3 axiomatic domains (propositional logic, arithmetic and group theory) demonstrate that our agent can bootstrap from only the axioms, self-improving in generating true and challenging conjectures and in finding proofs.