A long-standing goal of reinforcement learning is that algorithms can learn on training tasks and generalize well on unseen tasks like humans, where different tasks share similar dynamic with different reward functions. A general challenge is that it is nontrivial to quantitatively measure the similarities between these different tasks, which is vital for analyzing the task distribution and further designing algorithms with stronger generalization. To address this, we present a novel metric named Task Distribution Relevance (TDR) via optimal Q functions to capture the relevance of the task distribution quantitatively. In the case of tasks with a high TDR, i.e., the tasks differ significantly, we demonstrate that the Markovian policies cannot distinguish them, yielding poor performance accordingly. Based on this observation, we propose a framework of Reward Informed Dreamer (RID) with reward-informed world models, which captures invariant latent features over tasks and encodes reward signals into policies for distinguishing different tasks. In RID, we calculate the corresponding variational lower bound of the log-likelihood on the data, which includes a novel term to distinguish different tasks via states, based on reward-informed world models. Finally, extensive experiments in DeepMind control suite demonstrate that RID can significantly improve the performance of handling different tasks at the same time, especially for those with high TDR, and further generalize to unseen tasks effectively.
Importance sampling (IS) is a popular technique in off-policy evaluation, which re-weights the return of trajectories in the replay buffer to boost sample efficiency. However, training with IS can be unstable and previous attempts to address this issue mainly focus on analyzing the variance of IS. In this paper, we reveal that the instability is also related to a new notion of Reuse Bias of IS -- the bias in off-policy evaluation caused by the reuse of the replay buffer for evaluation and optimization. We theoretically show that the off-policy evaluation and optimization of the current policy with the data from the replay buffer result in an overestimation of the objective, which may cause an erroneous gradient update and degenerate the performance. We further provide a high-probability upper bound of the Reuse Bias, and show that controlling one term of the upper bound can control the Reuse Bias by introducing the concept of stability for off-policy algorithms. Based on these analyses, we finally present a novel Bias-Regularized Importance Sampling (BIRIS) framework along with practical algorithms, which can alleviate the negative impact of the Reuse Bias. Experimental results show that our BIRIS-based methods can significantly improve the sample efficiency on a series of continuous control tasks in MuJoCo.
Embodied agents in vision navigation coupled with deep neural networks have attracted increasing attention. However, deep neural networks are vulnerable to malicious adversarial noises, which may potentially cause catastrophic failures in Embodied Vision Navigation. Among these adversarial noises, universal adversarial perturbations (UAP), i.e., the image-agnostic perturbation applied on each frame received by the agent, are more critical for Embodied Vision Navigation since they are computation-efficient and application-practical during the attack. However, existing UAP methods do not consider the system dynamics of Embodied Vision Navigation. For extending UAP in the sequential decision setting, we formulate the disturbed environment under the universal noise $\delta$, as a $\delta$-disturbed Markov Decision Process ($\delta$-MDP). Based on the formulation, we analyze the properties of $\delta$-MDP and propose two novel Consistent Attack methods for attacking Embodied agents, which first consider the dynamic of the MDP by estimating the disturbed Q function and the disturbed distribution. In spite of victim models, our Consistent Attack can cause a significant drop in the performance for the Goalpoint task in habitat. Extensive experimental results indicate that there exist potential risks for applying Embodied Vision Navigation methods to the real world.
Though deep reinforcement learning (DRL) has obtained substantial success, it may encounter catastrophic failures due to the intrinsic uncertainty of both transition and observation. Most of the existing methods for safe reinforcement learning can only handle transition disturbance or observation disturbance since these two kinds of disturbance affect different parts of the agent; besides, the popular worst-case return may lead to overly pessimistic policies. To address these issues, we first theoretically prove that the performance degradation under transition disturbance and observation disturbance depends on a novel metric of Value Function Range (VFR), which corresponds to the gap in the value function between the best state and the worst state. Based on the analysis, we adopt conditional value-at-risk (CVaR) as an assessment of risk and propose a novel reinforcement learning algorithm of CVaR-Proximal-Policy-Optimization (CPPO) which formalizes the risk-sensitive constrained optimization problem by keeping its CVaR under a given threshold. Experimental results show that CPPO achieves a higher cumulative reward and is more robust against both observation and transition disturbances on a series of continuous control tasks in MuJoCo.
Recent works demonstrate that deep reinforcement learning (DRL) models are vulnerable to adversarial attacks which can decrease the victim's total reward by manipulating the observations. Compared with adversarial attacks in supervised learning, it is much more challenging to deceive a DRL model since the adversary has to infer the environmental dynamics. To address this issue, we reformulate the problem of adversarial attacks in function space and separate the previous gradient based attacks into several subspace. Following the analysis of the function space, we design a generic two-stage framework in the subspace where the adversary lures the agent to a target trajectory or a deceptive policy. In the first stage, we train a deceptive policy by hacking the environment, and discover a set of trajectories routing to the lowest reward. The adversary then misleads the victim to imitate the deceptive policy by perturbing the observations. Our method provides a tighter theoretical upper bound for the attacked agent's performance than the existing approaches. Extensive experiments demonstrate the superiority of our method and we achieve the state-of-the-art performance on both Atari and MuJoCo environments.