Hindsight Experience Replay Research Articles

From Semantics to Execution: Integrating Action Planning With Reinforcement Learning for Robotic Causal Problem-Solving.

Reinforcement learning is generally accepted to be an appropriate and successful method to learn robot control. Symbolic action planning is useful to resolve causal dependencies and to break a causally complex problem down into a sequence of simpler high-level actions. A problem with the integration of both approaches is that action planning is based on discrete high-level action- and state spaces, whereas reinforcement learning is usually driven by a continuous reward function. Recent advances in model-free reinforcement learning, specifically, universal value function approximators and hindsight experience replay, have focused on goal-independent methods based on sparse rewards that are only given at the end of a rollout, and only if the goal has been fully achieved. In this article, we build on these novel methods to facilitate the integration of action planning with model-free reinforcement learning. Specifically, the paper demonstrates how the reward-sparsity can serve as a bridge between the high-level and low-level state- and action spaces. As a result, we demonstrate that the integrated method is able to solve robotic tasks that involve non-trivial causal dependencies under noisy conditions, exploiting both data and knowledge.

Exploration with Multiple Random ε-Buffers in Off-Policy Deep Reinforcement Learning

In terms of deep reinforcement learning (RL), exploration is highly significant in achieving better generalization. In benchmark studies, ε-greedy random actions have been used to encourage exploration and prevent over-fitting, thereby improving generalization. Deep RL with random ε-greedy policies, such as deep Q-networks (DQNs), can demonstrate efficient exploration behavior. A random ε-greedy policy exploits additional replay buffers in an environment of sparse and binary rewards, such as in the real-time online detection of network securities by verifying whether the network is “normal or anomalous.” Prior studies have illustrated that a prioritized replay memory attributed to a complex temporal difference error provides superior theoretical results. However, another implementation illustrated that in certain environments, the prioritized replay memory is not superior to the randomly-selected buffers of random ε-greedy policy. Moreover, a key challenge of hindsight experience replay inspires our objective by using additional buffers corresponding to each different goal. Therefore, we attempt to exploit multiple random ε-greedy buffers to enhance explorations for a more near-perfect generalization with one original goal in off-policy RL. We demonstrate the benefit of off-policy learning from our method through an experimental comparison of DQN and a deep deterministic policy gradient in terms of discrete action, as well as continuous control for complete symmetric environments.

Deep Reinforcement Learning for the Navigation of Neurovascular Catheters

Abstract Endovascular catheters are necessary for state-ofthe- art treatments of life-threatening and time-critical diseases like strokes and heart attacks. Navigating them through the vascular tree is a highly challenging task. We present our preliminary results for the autonomous control of a guidewire through a vessel phantom with the help of Deep Reinforcement Learning. We trained Deep-Q-Network (DQN) and Deep Deterministic Policy Gradient (DDPG) agents on a simulated vessel phantom and evaluated the training performance. We also investigated the effect of the two enhancements Hindsight Experience Replay (HER) and Human Demonstration (HD) on the training speed of our agents. The results show that the agents are capable of learning to navigate a guidewire from a random start point in the vessel phantom to a random goal. This is achieved with an average success rate of 86.5% for DQN and 89.6% for DDPG. The use of HER and HD significantly increases the training speed. The results are promising and future research should address more complex vessel phantoms and the use of a combination of guidewire and catheter.

Goal-Oriented Dialogue Policy Learning from Failures

Reinforcement learning methods have been used for learning dialogue policies. However, learning an effective dialogue policy frequently requires prohibitively many conversations. This is partly because of the sparse rewards in dialogues, and the very few successful dialogues in early learning phase. Hindsight experience replay (HER) enables learning from failures, but the vanilla HER is inapplicable to dialogue learning due to the implicit goals. In this work, we develop two complex HER methods providing different tradeoffs between complexity and performance, and, for the first time, enabled HER-based dialogue policy learning. Experiments using a realistic user simulator show that our HER methods perform better than existing experience replay methods (as applied to deep Q-networks) in learning rate.

Guided goal generation for hindsight multi-goal reinforcement learning

Typical reinforcement learning (RL) can only perform a single task and thus cannot scale to problems for which an agent needs to perform multiple tasks, such as moving objects to different locations, which is relevant to real-world environments. Hindsight experience replay (HER) based on universal value functions shows promising results in such multi-goal settings by substituting achieved goals for the original goal, frequently giving the agent rewards. However, the achieved goals are limited to the current policy level and lack guidance for learning. We propose a novel guided goal-generation model for multi-goal RL named G-HER. Our method uses a conditional generative recurrent neural network (RNN) to explicitly model the relationship between policy level and goals, enabling the generation of various goals conditions on the different policy levels. Goals generated with a higher policy level provide better guidance for the RL agent, which is equivalent to using knowledge of successful policy in advance to guide the learning of current policy. Our model accelerates the generalization of substitute goals to the whole goal space. The G-HER algorithm is evaluated on several robotic manipulating tasks and demonstrates improved performance and sample efficiency.

Rewards Prediction-Based Credit Assignment for Reinforcement Learning With Sparse Binary Rewards

In reinforcement learning (RL), a reinforcement signal may be infrequent and delayed, not appearing immediately after the action that triggered the reward. To trace back what sequence of actions contributes to delayed rewards, e.g., credit assignment (CA), is one of the biggest challenges in RL. This challenge is aggravated under sparse binary rewards, especially when rewards are given only after successful completion of the task. To this end, a novel method consisting of key-action detection, among a sequence of actions to perform a task under sparse binary rewards, and CA strategy is proposed. The key-action defined as the most important action contributing to the reward is detected by a deep neural network that predicts future rewards based on the environment information. The rewards are re-assigned to the key-action and its adjacent actions, defined as adjacent-key-actions. Such re-assignment process enables increased success rate and convergence speed during training. For efficient re-assignment, two CA strategies are considered as part of proposed method. Proposed method is combined with hindsight experience replay (HER) for experiments in the OpenAI gym suite robotics environment. In the experiments, it is demonstrated that proposed method can detect key-actions and outperform the HER, increasing success rate and convergence speed, in the Fetch slide task, a type of task that is more exacting as compared to other tasks, but is addressed by few publications in the literature. From the experiments, a guideline for selecting CA strategy according to goal location is provided through goal distribution analysis with dot map.