Deep reinforcement learning is prone to overfitting, and traditional benchmarks such as Atari 2600 benchmark can exacerbate this problem. The Obstacle Tower Challenge addresses this by using randomized environments and separate seeds for training, validation, and test runs. This paper examines various improvements and best practices to the PPO algorithm using the Obstacle Tower Challenge to empirically study their impact with regards to generalization. Our experiments show that the combination provides state-of-the-art performance on the Obstacle Tower Challenge.

Among the data-efficient approaches for online adaptation in robotics (meta-learning, model-based reinforcement learning, etc.), repertoire-based learning (1) generates a large and diverse set policies in simulation that acts as a "reservoir" for future adaptations and (2) learns to pick online the best working policies according to the current situation (e.g., a damaged robot, a new object, etc.). Each of these policies performs a different task, for instance, walking in different directions; these policies are then sequenced with a planning algorithm to achieve the given task. In this paper, we relax the assumption of previous works that a single repertoire is enough for adaptation. Instead, we generate repertoires for many different situations (e.g., with a missing leg, on different floors, etc.) in simulation that act as priors for adaptation. Our main contribution is an algorithm, APROL (Adaptive Prior selection for Repertoire-based Online Learning) to plan the next action by incorporating these priors when the robot has no information about the current situation. We evaluate APROL on two simulated tasks: (1) pushing unknown objects of various shapes and sizes with a kuka arm and (2) a goal reaching task with a damaged hexapod robot. We compare with "Reset-free Trial and Error" (RTE) and various single repertoire-based baselines. The results show that APROL solves both tasks in less interaction time than the baselines. Additionally, we demonstrate APROL on a real, damaged hexapod that quickly learns compensatory policies to reach a goal by avoiding obstacle in the path.

We present a training pipeline for the autonomous driving task given the current camera image and vehicle speed as the input to produce the throttle, brake, and steering control output. The simulator Airsim's convenient weather and lighting API provides a sufficient diversity during training which can be very helpful to increase the trained policy's robustness. In order to not limit the possible policy's performance, we use a continuous and deterministic control policy setting. We utilize ResNet-34 as our actor and critic networks with some slight changes in the fully connected layers. Considering human's mastery of this task and the high-complexity nature of this task, we first use imitation learning to mimic the given human policy and leverage the trained policy and its weights to the reinforcement learning phase for which we use DDPG. This combination shows a considerable performance boost comparing to both pure imitation learning and pure DDPG for the autonomous driving task.

Despite the tremendous advances that have been made in the last decade on developing useful machine-learning applications, their wider adoption has been hindered by the lack of strong assurance guarantees that can be made about their behavior. In this paper, we consider how formal verification techniques developed for traditional software systems can be repurposed for verification of reinforcement learning-enabled ones, a particularly important class of machine learning systems. Rather than enforcing safety by examining and altering the structure of a complex neural network implementation, our technique uses blackbox methods to synthesizes deterministic programs, simpler, more interpretable, approximations of the network that can nonetheless guarantee desired safety properties are preserved, even when the network is deployed in unanticipated or previously unobserved environments. Our methodology frames the problem of neural network verification in terms of a counterexample and syntax-guided inductive synthesis procedure over these programs. The synthesis procedure searches for both a deterministic program and an inductive invariant over an infinite state transition system that represents a specification of an application's control logic. Additional specifications defining environment-based constraints can also be provided to further refine the search space. Synthesized programs deployed in conjunction with a neural network implementation dynamically enforce safety conditions by monitoring and preventing potentially unsafe actions proposed by neural policies. Experimental results over a wide range of cyber-physical applications demonstrate that software-inspired formal verification techniques can be used to realize trustworthy reinforcement learning systems with low overhead.

The aim of multi-task reinforcement learning is two-fold: (1) efficiently learn by training against multiple tasks and (2) quickly adapt, using limited samples, to a variety of new tasks. In this work, the tasks correspond to reward functions for environments with the same (or similar) dynamical models. We propose to learn a dynamical model during the training process and use this model to perform sample-efficient adaptation to new tasks at test time. We use significantly fewer samples by performing policy optimization only in a "virtual" environment whose transitions are given by our learned dynamical model. Our algorithm sequentially trains against several tasks. Upon encountering a new task, we first warm-up a policy on our learned dynamical model, which requires no new samples from the environment. We then adapt the dynamical model with samples from this policy in the real environment. We evaluate our approach on several continuous control benchmarks and demonstrate its efficacy over MAML, a state-of-the-art meta-learning algorithm, on these tasks.

Instability and slowness are two main problems in deep reinforcement learning. Even if proximal policy optimization is the state of the art, it still suffers from these two problems. We introduce an improved algorithm based on proximal policy optimization (PPO), mixed distributed proximal policy optimization (MDPPO), and show that it can accelerate and stabilize the training process. In our algorithm, multiple different policies train simultaneously and each of them controls several identical agents that interact with environments. Actions are sampled by each policy separately as usual but the trajectories for training process are collected from all agents, instead of only one policy. We find that if we choose some auxiliary trajectories elaborately to train policies, the algorithm will be more stable and quicker to converge especially in the environments with sparse rewards.

Replay memory plays an important role in stable learning and fast convergence of deep reinforcement learning algorithms [1] that are methods of approximating a value or a policy function using deep neural networks [2]. The study of replay memory in reinforcement learning started from [3] and played a major role in training reinforcement learning agents to play Atari 2600 games with a Deep Q-Network (DQN) [4]. In addition, replay memory is used in other off-policy reinforcement learning algorithms such as DDPG [5] and ACER [6]. In [7], after analyzing the importance of the data in the replay memory, a probability distribution is assigned to enable efficient learning through prioritization based on the Figure 1: Proposed dual memory structure.

Let's say I am given an Excel sheet with data about various fruits and I have to tell which look like Apples. What I will do is ask a question "Which fruits are red and round?" and divide all fruits which answer yes and no to the question. Now, All Red and Round fruits might not be apples and all apples won't be red and round. So I will ask a question "Which fruits have red or yellow colour hints on them? " on red and round fruits and will ask "Which fruits are green and round?" on not red and round fruits. Based on these questions I can tell with considerable accuracy which are apples. This cascade of questions is what a decision tree is. However, this is a decision tree based on my intuition.

Compared with the classical policy gradient algorithm 1992), actor-critic tracks the action-value function (critic) in policy gradient in an online(Williams, manner, and alternatively updates the policy (actor) and the critic. On the one hand, the online update of critic significantly reduces the variance of policy gradient and hence leads to faster convergence. On the other hand, it also introduces algorithmic instability, which is often observed in practice (Islam et al., 2017) and parallels the notoriously unstable training of generative adversarial and Vinyals, 2016). Such instability of actor-critic originates from severalnetworks (Pfau intertwining challenges, including(i) function approximation of actor and critic, (ii) improper choice of stepsizes, (iii) the noise arising from stochastic approximation, (iv) the asynchrony between actor and critic, and (v) possibly off-policy data used in the update of critic. As a result, the convergence of actor-critic remains much less well understood than that of policy gradient, which itself is open. Consequently, the practical use of actor-critic often lacks theoretical guidance. In this paper, we aim to theoretically understand the algorithmic instability of actor-critic.

We study two time-scale linear stochastic approximation algorithms, which can be used to model well-known reinforcement learning algorithms such as GTD, GTD2, and TDC. We present finite-time performance bounds for the case where the learning rate is fixed. The key idea in obtaining these bounds is to use a Lyapunov function motivated by singular perturbation theory for linear differential equations. We use the bound to design an adaptive learning rate scheme which significantly improves the convergence rate over the known optimal polynomial decay rule in our experiments, and can be used to potentially improve the performance of any other schedule where the learning rate is changed at pre-determined time instants.