The use of human demonstrations in reinforcement learning has proven to significantly improve agent performance. However, any requirement for a human to manually 'teach' the model is somewhat antithetical to the goals of reinforcement learning. This paper attempts to minimize human involvement in the learning process while retaining the performance advantages by using a single human example collected through a simple-to-use virtual reality simulation to assist with RL training. Our method augments a single demonstration to generate numerous human-like demonstrations that, when combined with Deep Deterministic Policy Gradients and Hindsight Experience Replay (DDPG + HER) significantly improve training time on simple tasks and allows the agent to solve a complex task (block stacking) that DDPG + HER alone cannot solve. The model achieves this significant training advantage using a single human example, requiring less than a minute of human input. Moreover, despite learning from a human example, the agent is not constrained to human-level performance, often learning a policy that is significantly different from the human demonstration.
相關內容
Large Language Models (LLMs), such as the Generative Pretrained Transformer (GPT), have achieved tremendous success in various language tasks, but their emergent abilities have also raised many questions, concerns, and challenges that need to be addressed. To gain a better understanding of the models' inner mechanisms, we analyze the hidden state and channel wave dynamics in a small GPT, focusing on the coherence of wave patterns in terms of cross-channel correlation and individual auto-correlation. Our findings suggest that wave dynamics offer consistent and repeatable intrinsic oscillation modes, along with context-aware plasticity and expressiveness in language generation. By analyzing wave patterns, coherence, and clustering, we provide a systematic way to identify and interpret the functionality of the hidden state channels, paving the way to understand and control higher-level language pattern formation. In addition, we investigate the Poisson statistics of spelling errors in text sequence generation across various levels of model training and observe a phase-transition-like process. As coherence builds up, there is a competition between the generation of correct and misspelled words. However, once the model is adequately trained and significant coherence has emerged, the coherent process becomes strong enough to effectively suppress spelling errors, preventing the cascade amplification of defects. The distribution of correct spellings transitions from Poissonian to Sub-Poissonian, while the distribution of misspellings shows the opposite trend. By leveraging concepts and techniques from quantum physics, we gain novel insights into the dynamics of the small GPT. This approach can be extended to larger language models that exhibit more complex coherent language patterns, opening up opportunities to interpret their emergent capabilities and develop more specialized models.
Deploying reinforcement learning agents in the real world can be challenging due to the risks associated with learning through trial and error. We propose a task-agnostic method that leverages small sets of safe and unsafe demonstrations to improve the safety of RL agents during learning. The method compares the current trajectory of the agent with both sets of demonstrations at every step, and filters the trajectory if it resembles the unsafe demonstrations. We perform ablation studies on different filtering strategies and investigate the impact of the number of demonstrations on performance. Our method is compatible with any stand-alone RL algorithm and can be applied to any task. We evaluate our method on three tasks from OpenAI Gym's Mujoco benchmark and two state-of-the-art RL algorithms. The results demonstrate that our method significantly reduces the crash rate of the agent while converging to, and in most cases even improving, the performance of the stand-alone agent.
When autonomous vehicles are deployed on public roads, they will encounter countless and diverse driving situations. Many manually designed driving policies are difficult to scale to the real world. Fortunately, reinforcement learning has shown great success in many tasks by automatic trial and error. However, when it comes to autonomous driving in interactive dense traffic, RL agents either fail to learn reasonable performance or necessitate a large amount of data. Our insight is that when humans learn to drive, they will 1) make decisions over the high-level skill space instead of the low-level control space and 2) leverage expert prior knowledge rather than learning from scratch. Inspired by this, we propose ASAP-RL, an efficient reinforcement learning algorithm for autonomous driving that simultaneously leverages motion skills and expert priors. We first parameterized motion skills, which are diverse enough to cover various complex driving scenarios and situations. A skill parameter inverse recovery method is proposed to convert expert demonstrations from control space to skill space. A simple but effective double initialization technique is proposed to leverage expert priors while bypassing the issue of expert suboptimality and early performance degradation. We validate our proposed method on interactive dense-traffic driving tasks given simple and sparse rewards. Experimental results show that our method can lead to higher learning efficiency and better driving performance relative to previous methods that exploit skills and priors differently. Code is open-sourced to facilitate further research.
In Reinforcement Learning (RL), an agent acts in an unknown environment to maximize the expected cumulative discounted sum of an external reward signal, i.e., the expected return. In practice, in many tasks of interest, such as policy optimization, the agent usually spends its interaction budget by collecting episodes of fixed length within a simulator (i.e., Monte Carlo simulation). However, given the discounted nature of the RL objective, this data collection strategy might not be the best option. Indeed, the rewards taken in early simulation steps weigh exponentially more than future rewards. Taking a cue from this intuition, in this paper, we design an a-priori budget allocation strategy that leads to the collection of trajectories of different lengths, i.e., truncated. The proposed approach provably minimizes the width of the confidence intervals around the empirical estimates of the expected return of a policy. After discussing the theoretical properties of our method, we make use of our trajectory truncation mechanism to extend Policy Optimization via Importance Sampling (POIS, Metelli et al., 2018) algorithm. Finally, we conduct a numerical comparison between our algorithm and POIS: the results are consistent with our theory and show that an appropriate truncation of the trajectories can succeed in improving performance.
In this paper, we propose a novel data-driven approach for learning and control of quadrotor UAVs based on the Koopman operator and extended dynamic mode decomposition (EDMD). Building observables for EDMD based on conventional methods like Euler angles or quaternions to represent orientation is known to involve singularities. To address this issue, we employ a set of physics-informed observables based on the underlying topology of the nonlinear system. We use rotation matrices to directly represent the orientation dynamics and obtain a lifted linear representation of the nonlinear quadrotor dynamics in the SE(3) manifold. This EDMD model leads to accurate prediction and can generalize to several validation sets. Further, we design a linear model predictive controller (MPC) based on the proposed EDMD model to track agile reference trajectories. Simulation results show that the proposed MPC controller can run as fast as 100 Hz and is able to track arbitrary reference trajectories with good accuracy. Implementation details can be found in \url{//github.com/sriram-2502/KoopmanMPC_Quadrotor}
The past few years have seen rapid progress in combining reinforcement learning (RL) with deep learning. Various breakthroughs ranging from games to robotics have spurred the interest in designing sophisticated RL algorithms and systems. However, the prevailing workflow in RL is to learn tabula rasa, which may incur computational inefficiency. This precludes continuous deployment of RL algorithms and potentially excludes researchers without large-scale computing resources. In many other areas of machine learning, the pretraining paradigm has shown to be effective in acquiring transferable knowledge, which can be utilized for a variety of downstream tasks. Recently, we saw a surge of interest in Pretraining for Deep RL with promising results. However, much of the research has been based on different experimental settings. Due to the nature of RL, pretraining in this field is faced with unique challenges and hence requires new design principles. In this survey, we seek to systematically review existing works in pretraining for deep reinforcement learning, provide a taxonomy of these methods, discuss each sub-field, and bring attention to open problems and future directions.
The development of autonomous agents which can interact with other agents to accomplish a given task is a core area of research in artificial intelligence and machine learning. Towards this goal, the Autonomous Agents Research Group develops novel machine learning algorithms for autonomous systems control, with a specific focus on deep reinforcement learning and multi-agent reinforcement learning. Research problems include scalable learning of coordinated agent policies and inter-agent communication; reasoning about the behaviours, goals, and composition of other agents from limited observations; and sample-efficient learning based on intrinsic motivation, curriculum learning, causal inference, and representation learning. This article provides a broad overview of the ongoing research portfolio of the group and discusses open problems for future directions.
We introduce DeepNash, an autonomous agent capable of learning to play the imperfect information game Stratego from scratch, up to a human expert level. Stratego is one of the few iconic board games that Artificial Intelligence (AI) has not yet mastered. This popular game has an enormous game tree on the order of $10^{535}$ nodes, i.e., $10^{175}$ times larger than that of Go. It has the additional complexity of requiring decision-making under imperfect information, similar to Texas hold'em poker, which has a significantly smaller game tree (on the order of $10^{164}$ nodes). Decisions in Stratego are made over a large number of discrete actions with no obvious link between action and outcome. Episodes are long, with often hundreds of moves before a player wins, and situations in Stratego can not easily be broken down into manageably-sized sub-problems as in poker. For these reasons, Stratego has been a grand challenge for the field of AI for decades, and existing AI methods barely reach an amateur level of play. DeepNash uses a game-theoretic, model-free deep reinforcement learning method, without search, that learns to master Stratego via self-play. The Regularised Nash Dynamics (R-NaD) algorithm, a key component of DeepNash, converges to an approximate Nash equilibrium, instead of 'cycling' around it, by directly modifying the underlying multi-agent learning dynamics. DeepNash beats existing state-of-the-art AI methods in Stratego and achieved a yearly (2022) and all-time top-3 rank on the Gravon games platform, competing with human expert players.
Contrastive learning models have achieved great success in unsupervised visual representation learning, which maximize the similarities between feature representations of different views of the same image, while minimize the similarities between feature representations of views of different images. In text summarization, the output summary is a shorter form of the input document and they have similar meanings. In this paper, we propose a contrastive learning model for supervised abstractive text summarization, where we view a document, its gold summary and its model generated summaries as different views of the same mean representation and maximize the similarities between them during training. We improve over a strong sequence-to-sequence text generation model (i.e., BART) on three different summarization datasets. Human evaluation also shows that our model achieves better faithfulness ratings compared to its counterpart without contrastive objectives.
This paper presents a new multi-objective deep reinforcement learning (MODRL) framework based on deep Q-networks. We propose the use of linear and non-linear methods to develop the MODRL framework that includes both single-policy and multi-policy strategies. The experimental results on two benchmark problems including the two-objective deep sea treasure environment and the three-objective mountain car problem indicate that the proposed framework is able to converge to the optimal Pareto solutions effectively. The proposed framework is generic, which allows implementation of different deep reinforcement learning algorithms in different complex environments. This therefore overcomes many difficulties involved with standard multi-objective reinforcement learning (MORL) methods existing in the current literature. The framework creates a platform as a testbed environment to develop methods for solving various problems associated with the current MORL. Details of the framework implementation can be referred to //www.deakin.edu.au/~thanhthi/drl.htm.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191