We consider the off-policy evaluation problem of reinforcement learning using deep convolutional neural networks. We analyze the deep fitted Q-evaluation method for estimating the expected cumulative reward of a target policy, when the data are generated from an unknown behavior policy. We show that, by choosing network size appropriately, one can leverage any low-dimensional manifold structure in the Markov decision process and obtain a sample-efficient estimator without suffering from the curse of high data ambient dimensionality. Specifically, we establish a sharp error bound for fitted Q-evaluation, which depends on the intrinsic dimension of the state-action space, the smoothness of Bellman operator, and a function class-restricted $\chi^2$-divergence. It is noteworthy that the restricted $\chi^2$-divergence measures the behavior and target policies' {\it mismatch in the function space}, which can be small even if the two policies are not close to each other in their tabular forms. We also develop a novel approximation result for convolutional neural networks in Q-function estimation. Numerical experiments are provided to support our theoretical analysis.

There are three types of common machine learning approaches: 1) supervised learning, where a learning system learns a latent map based on labeled examples, 2) unsupervised learning, where a learning system establishes a model for data distribution based on unlabeled examples, and 3) Reinforcement Learning, where a decision-making system is trained to make optimal decisions. From the designer's point-of-view, all kinds of learning are supervised by a loss function. The sources of supervision must be defined by humans. One way to do this is by the loss function. In supervised learning, the ground truth label is provided.

This means that evaluating and playing around with different algorithms is easy. Of course you can extend keras-rl according to your own needs. You can use built-in Keras callbacks and metrics or define your own. Even more so, it is easy to implement your own environments and even algorithms by simply extending some simple abstract classes. You can find more information on each agent in the doc. If you want to run the examples, you'll also have to install: This is a very simple example and it should converge relatively quickly, so it's a great way to get started!

We address the problem of safe reinforcement learning from pixel observations. Inherent challenges in such settings are (1) a trade-off between reward optimization and adhering to safety constraints, (2) partial observability, and (3) high-dimensional observations. We formalize the problem in a constrained, partially observable Markov decision process framework, where an agent obtains distinct reward and safety signals. To address the curse of dimensionality, we employ a novel safety critic using the stochastic latent actor-critic (SLAC) approach. The latent variable model predicts rewards and safety violations, and we use the safety critic to train safe policies. Using well-known benchmark environments, we demonstrate competitive performance over existing approaches with respects to computational requirements, final reward return, and satisfying the safety constraints.

The capabilities of a robot will be increased significantly by exploiting throwing behavior. In particular, throwing will enable robots to rapidly place the object into the target basket, located outside its feasible kinematic space, without traveling to the desired location. In previous approaches, the robot often learned a parameterized throwing kernel through analytical approaches, imitation learning, or hand-coding. There are many situations in which such approaches do not work/generalize well due to various object shapes, heterogeneous mass distribution, and also obstacles that might be presented in the environment. It is obvious that a method is needed to modulate the throwing kernel through its meta parameters. In this paper, we tackle object throwing problem through a deep reinforcement learning approach that enables robots to precisely throw objects into moving baskets while there are obstacles obstructing the path. To the best of our knowledge, we are the first group that addresses throwing objects with obstacle avoidance. Such a throwing skill not only increases the physical reachability of a robot arm but also improves the execution time. In particular, the robot detects the pose of the target object, basket, and obstacle at each time step, predicts the proper grasp configuration for the target object, and then infers appropriate parameters to throw the object into the basket. Due to safety constraints, we develop a simulation environment in Gazebo to train the robot and then use the learned policy in real-robot directly. To assess the performers of the proposed approach, we perform extensive sets of experiments in both simulation and real robots in three scenarios. Experimental results showed that the robot could precisely throw a target object into the basket outside its kinematic range and generalize well to new locations and objects without colliding with obstacles.

ABSTRACT With the emerging connected-vehicle technologies and smart roads, the need for intelligent adaptive traffic signal controls is more than ever before. This paper proposes a novel Economicdriven Adaptive Traffic Signal Control (eATSC) model with a hyper control variable - interest rate defined in economics for traffic signal control at signalized intersections. The eATSC uses a continuous compounding function that captures both the total number of vehicles and the accumulated waiting time of each vehicle to compute penalties for different directions. The computed penalties grow with waiting time and is used for signal control decisions. Each intersection is assigned two intelligent agents adjusting interest rate and signal length for different directions according to the traffic patterns, respectively. The problem is formulated as a Markov Decision Process (MDP) problem to reduce congestions, and a two-agent Double Dueling Deep Q Network (DDDQN) is utilized to solve the problem. Under the optimal policy, the agents can select the optimal interest rates and signal time to minimize the likelihood of traffic congestions. To evaluate the superiority of our method, a VISSIM simulation model with classic four-leg signalized intersections is developed. The results indicate that the proposed model is adequately able to maintain healthy traffic flow at the intersection. INTRODUCTION Many studies have shown that adaptive signal control (ASC) improves traffic performance, such as emissions, travel time, and fuel consumption by at least 10% [1].

In continuous control, exploration is often performed through undirected strategies in which parameters of the networks or selected actions are perturbed by random noise. Although the deep setting of undirected exploration has been shown to improve the performance of on-policy methods, they introduce an excessive computational complexity and are known to fail in the off-policy setting. The intrinsically motivated exploration is an effective alternative to the undirected strategies, but they are usually studied for discrete action domains. In this paper, we investigate how intrinsic motivation can effectively be combined with deep reinforcement learning in the control of continuous systems to obtain a directed exploratory behavior. We adapt the existing theories on animal motivational systems into the reinforcement learning paradigm and introduce a novel and scalable directed exploration strategy. The introduced approach, motivated by the maximization of the value function's error, can benefit from a collected set of experiences by extracting useful information and unify the intrinsic exploration motivations in the literature under a single exploration objective. An extensive set of empirical studies demonstrate that our framework extends to larger and more diverse state spaces, dramatically improves the baselines, and outperforms the undirected strategies significantly.

In this paper, we study the problem of coverage of an environment with an energy-constrained robot in the presence of multiple charging stations. As the robot's on-board power supply is limited, it might not have enough energy to cover all the points in the environment with a single charge. Instead, it will need to stop at one or more charging stations to recharge its battery intermittently. The robot cannot violate the energy constraint, i.e., visit a location with negative available energy. To solve this problem, we propose a deep Q-learning framework that produces a policy to maximize the coverage and minimize the budget violations. Our proposed framework also leverages the memory of a recurrent neural network (RNN) to better suit this multi-objective optimization problem. We have tested the presented framework within a 16 x 16 grid environment having charging stations and various obstacle configurations. Results show that our proposed method finds feasible solutions and outperforms a comparable existing technique.

Guided exploration with expert demonstrations improves data efficiency for reinforcement learning, but current algorithms often overuse expert information. We propose a novel algorithm to speed up Q-learning with the help of a limited amount of imperfect expert demonstrations. The algorithm avoids excessive reliance on expert data by relaxing the optimal expert assumption and gradually reducing the usage of uninformative expert data. Experimentally, we evaluate our approach on a sparse-reward chain environment and six more complicated Atari games with delayed rewards. With the proposed methods, we can achieve better results than Deep Q-learning from Demonstrations (Hester et al., 2017) in most environments.

Actor-critic (AC) algorithms are a class of model-free deep reinforcement learning algorithms, which have proven their efficacy in diverse domains, especially in solving continuous control problems. Improvement of exploration (action entropy) and exploitation (expected return) using more efficient samples is a critical issue in AC algorithms. A basic strategy of a learning algorithm is to facilitate indiscriminately exploring all of the environment state space, as well as to encourage exploring rarely visited states rather than frequently visited one. Under this strategy, we propose a new method to boost exploration through an intrinsic reward, based on measurement of a state's novelty and the associated benefit of exploring the state (with regards to policy optimization), altogether called plausible novelty. With incentivized exploration of plausible novel states, an AC algorithm is able to improve its sample efficiency and hence training performance. The new method is verified by extensive simulations of continuous control tasks of MuJoCo environments on a variety of prominent off-policy AC algorithms.