However, much of the successes reinforces the fact that much of the explored trajectories have also been attributed to well specified reward functions do not actually participate in the reward learning process which ground the agent's behavior and subsequent task in and in fact their best change to affect the reward function the expected manner. As prior works have argued, the specification is once they get sampled and queried to the human in the of low level reward functions for seemingly easy loop. We posit that this untapped data source can greatly tasks could be quite difficult and may still result in inexplicable improve reward recovery and reduce the feedback sample and unexpected results (Verma et al. 2019, 2021; complexity. Our second observation notes that the reward Gopalakrishnan, Verma, and Kambhampati 2021a,b) potentially function being learnt may not conform to the structure of affecting trust between Human-AI (Zahedi et al. 2021, the state space simply because it doesn't get exposed to as 2022). For example, works like (Krakovna et al. 2020; Vamplew many data points (in comparison to, say, the policy approximation et al. 2018) have raised the issues of reward hacking function). We utilize our observations to improve and reward exploitation where the RL agents would discover performance of RL agents in recovering the underlying behaviors that seems to be "cheating" or incorrect and reward function and learn a good policy by exploiting the yet maximize the expected cumulative reward. This has also rich unlabeled trajectory data. Although works like SURF gotten attention from the explainable AI community where (Park et al. 2022) have proposed a semi-supervised learning they attempt to analyze whether the agent is actually behaving approach to utilize unlabeled trajectory data, they would in the intended manner (Verma, Kharkwal, and Kambhampati generate labels for unlabeled trajectories and use these 2022; Sreedharan et al. 2020; Kambhampati et al. data points as if they were given by the human in the loop.

Offline goal-conditioned reinforcement learning (GCRL) can be challenging due to overfitting to the given dataset. To generalize agents' skills outside the given dataset, we propose a goal-swapping procedure that generates additional trajectories. To alleviate the problem of noise and extrapolation errors, we present a general offline reinforcement learning method called deterministic Q-advantage policy gradient (DQAPG). In the experiments, DQAPG outperforms state-of-the-art goal-conditioned offline RL methods in a wide range of benchmark tasks, and goal-swapping further improves the test results. It is noteworthy, that the proposed method obtains good performance on the challenging dexterous in-hand manipulation tasks for which the prior methods failed.

In reinforcement learning (RL), it is easier to solve a task if given a good representation. While deep RL should automatically acquire such good representations, prior work often finds that learning representations in an end-to-end fashion is unstable and instead equip RL algorithms with additional representation learning parts (e.g., auxiliary losses, data augmentation). How can we design RL algorithms that directly acquire good representations? In this paper, instead of adding representation learning parts to an existing RL algorithm, we show (contrastive) representation learning methods can be cast as RL algorithms in their own right. To do this, we build upon prior work and apply contrastive representation learning to action-labeled trajectories, in such a way that the (inner product of) learned representations exactly corresponds to a goal-conditioned value function. We use this idea to reinterpret a prior RL method as performing contrastive learning, and then use the idea to propose a much simpler method that achieves similar performance. Across a range of goal-conditioned RL tasks, we demonstrate that contrastive RL methods achieve higher success rates than prior non-contrastive methods, including in the offline RL setting. We also show that contrastive RL outperforms prior methods on image-based tasks, without using data augmentation or auxiliary objectives.

Prior work has proposed a simple strategy for reinforcement learning (RL): label experience with the outcomes achieved in that experience, and then imitate the relabeled experience. These outcome-conditioned imitation learning methods are appealing because of their simplicity, strong performance, and close ties with supervised learning. However, it remains unclear how these methods relate to the standard RL objective, reward maximization. In this paper, we formally relate outcome-conditioned imitation learning to reward maximization, drawing a precise relationship between the learned policy and Q-values and explaining the close connections between these methods and prior EM-based policy search methods. This analysis shows that existing outcome-conditioned imitation learning methods do not necessarily improve the policy, but a simple modification results in a method that does guarantee policy improvement, under some assumptions.

Scientists have taken a key step toward harnessing a form of artificial intelligence known as deep reinforcement learning, or DRL, to protect computer networks. When faced with sophisticated cyberattacks in a rigorous simulation setting, deep reinforcement learning was effective at stopping adversaries from reaching their goals up to 95 percent of the time. The outcome offers promise for a role for autonomous AI in proactive cyber defense. Scientists from the Department of Energy's Pacific Northwest National Laboratory documented their findings in a research paper and presented their work Feb. 14 at a workshop on AI for Cybersecurity during the annual meeting of the Association for the Advancement of Artificial Intelligence in Washington, D.C. The starting point was the development of a simulation environment to test multistage attack scenarios involving distinct types of adversaries.

Reinforcement Learning (RL) and Imitation Learning (IL) have made great progress in robotic control in recent years. However, these methods show obvious deterioration for new tasks that need to be completed through new combinations of actions. RL methods heavily rely on reward functions that cannot generalize well for new tasks, while IL methods are limited by expert demonstrations which do not cover new tasks. In contrast, humans can easily complete these tasks with the fragmented knowledge learned from task-agnostic experience. Inspired by this observation, this paper proposes a task-agnostic learning method (TAL for short) that can learn fragmented knowledge from task-agnostic data to accomplish new tasks. TAL consists of four stages. First, the task-agnostic exploration is performed to collect data from interactions with the environment. The collected data is organized via a knowledge graph. Compared with the previous sequential structure, the knowledge graph representation is more compact and fits better for environment exploration. Second, an action feature extractor is proposed and trained using the collected knowledge graph data for task-agnostic fragmented knowledge learning. Third, a candidate action generator is designed, which applies the action feature extractor on a new task to generate multiple candidate action sets. Finally, an action proposal is designed to produce the probabilities for actions in a new task according to the environmental information. The probabilities are then used to select actions to be executed from multiple candidate action sets to form the plan. Experiments on a virtual indoor scene show that the proposed method outperforms the state-of-the-art offline RL method: CQL by 35.28% and the IL method: BC by 22.22%.

Reinforcement learning (RL) and trajectory optimization (TO) present strong complementary advantages. On one hand, RL approaches are able to learn global control policies directly from data, but generally require large sample sizes to properly converge towards feasible policies. On the other hand, TO methods are able to exploit gradient-based information extracted from simulators to quickly converge towards a locally optimal control trajectory which is only valid within the vicinity of the solution. Over the past decade, several approaches have aimed to adequately combine the two classes of methods in order to obtain the best of both worlds. Following on from this line of research, we propose several improvements on top of these approaches to learn global control policies quicker, notably by leveraging sensitivity information stemming from TO methods via Sobolev learning, and augmented Lagrangian techniques to enforce the consensus between TO and policy learning. We evaluate the benefits of these improvements on various classical tasks in robotics through comparison with existing approaches in the literature.

How have individuals of social animals in nature evolved to learn from each other, and what would be the optimal strategy for such learning in a specific environment? Here, we address both problems by employing a deep reinforcement learning model to optimize the social learning strategies (SLSs) of agents in a cooperative game in a multi-dimensional landscape. Throughout the training for maximizing the overall payoff, we find that the agent spontaneously learns various concepts of social learning, such as copying, focusing on frequent and well-performing neighbors, self-comparison, and the importance of balancing between individual and social learning, without any explicit guidance or prior knowledge about the system. The SLS from a fully trained agent outperforms all of the traditional, baseline SLSs in terms of mean payoff. We demonstrate the superior performance of the reinforcement learning agent in various environments, including temporally changing environments and real social networks, which also verifies the adaptability of our framework to different social settings.

Large-scale training have propelled significant progress in various sub-fields of AI such as computer vision and natural language processing. However, building robot learning systems at a comparable scale remains challenging. To develop robots that can perform a wide range of skills and adapt to new scenarios, efficient methods for collecting vast and diverse amounts of data on physical robot systems are required, as well as the capability to train high-capacity policies using such datasets. In this work, we propose a framework for scaling robot learning, with specific focus on multi-task and multi-scene manipulation in kitchen environments, both in simulation and in the real world. Our proposed framework, CACTI, comprises four stages that separately handle: data collection, data augmentation, visual representation learning, and imitation policy training, to enable scalability in robot learning . We make use of state-of-the-art generative models as part of the data augmentation stage, and use pre-trained out-of-domain visual representations to improve training efficiency. Experimental results demonstrate the effectiveness of our approach. On a real robot setup, CACTI enables efficient training of a single policy that can perform 10 manipulation tasks involving kitchen objects, and is robust to varying layouts of distractors. In a simulated kitchen environment, CACTI trains a single policy to perform 18 semantic tasks across 100 layout variations for each individual task. We will release the simulation task benchmark and augmented datasets in both real and simulated environments to facilitate future research.

This paper is concerned with offline reinforcement learning (RL), which learns using pre-collected data without further exploration. Effective offline RL would be able to accommodate distribution shift and limited data coverage. However, prior algorithms or analyses either suffer from suboptimal sample complexities or incur high burn-in cost to reach sample optimality, thus posing an impediment to efficient offline RL in sample-starved applications. We demonstrate that the model-based (or "plug-in") approach achieves minimax-optimal sample complexity without burn-in cost for tabular Markov decision processes (MDPs). Concretely, consider a finite-horizon (resp. $\gamma$-discounted infinite-horizon) MDP with $S$ states and horizon $H$ (resp. effective horizon $\frac{1}{1-\gamma}$), and suppose the distribution shift of data is reflected by some single-policy clipped concentrability coefficient $C^{\star}_{\text{clipped}}$. We prove that model-based offline RL yields $\varepsilon$-accuracy with a sample complexity of \[ \begin{cases} \frac{H^{4}SC_{\text{clipped}}^{\star}}{\varepsilon^{2}} & (\text{finite-horizon MDPs}) \frac{SC_{\text{clipped}}^{\star}}{(1-\gamma)^{3}\varepsilon^{2}} & (\text{infinite-horizon MDPs}) \end{cases} \] up to log factor, which is minimax optimal for the entire $\varepsilon$-range. The proposed algorithms are ``pessimistic'' variants of value iteration with Bernstein-style penalties, and do not require sophisticated variance reduction. Our analysis framework is established upon delicate leave-one-out decoupling arguments in conjunction with careful self-bounding techniques tailored to MDPs.