Reinforcement learning (RL) has had many successes in both "deep" and "shallow" settings. In both cases, significant hyperparameter tuning is often required to achieve good performance. Furthermore, when nonlinear function approximation is used, non-stationarity in the state representation can lead to learning instability. A variety of techniques exist to combat this --- most notably large experience replay buffers or the use of multiple parallel actors. These techniques come at the cost of moving away from the online RL problem as it is traditionally formulated (i.e., a single agent learning online without maintaining a large database of training examples). Meta-learning can potentially help with both these issues by tuning hyperparameters online and allowing the algorithm to more robustly adjust to non-stationarity in a problem. This paper applies meta-gradient descent to derive a set of step-size tuning algorithms specifically for online RL control with eligibility traces. Our novel technique, Metatrace, makes use of an eligibility trace analogous to methods like $TD(\lambda)$. We explore tuning both a single scalar step-size and a separate step-size for each learned parameter. We evaluate Metatrace first for control with linear function approximation in the classic mountain car problem and then in a noisy, non-stationary version. Finally, we apply Metatrace for control with nonlinear function approximation in 5 games in the Arcade Learning Environment where we explore how it impacts learning speed and robustness to initial step-size choice. Results show that the meta-step-size parameter of Metatrace is easy to set, Metatrace can speed learning, and Metatrace can allow an RL algorithm to deal with non-stationarity in the learning task.

Coordinating agents to complete a set of tasks with intercoupled temporal and resource constraints is computationally challenging, yet human domain experts can solve these difficult scheduling problems using paradigms learned through years of apprenticeship. A process for manually codifying this domain knowledge within a computational framework is necessary to scale beyond the ``single-expert, single-trainee" apprenticeship model. However, human domain experts often have difficulty describing their decision-making processes, causing the codification of this knowledge to become laborious. We propose a new approach for capturing domain-expert heuristics through a pairwise ranking formulation. Our approach is model-free and does not require enumerating or iterating through a large state space. We empirically demonstrate that this approach accurately learns multifaceted heuristics on a synthetic data set incorporating job-shop scheduling and vehicle routing problems, as well as on two real-world data sets consisting of demonstrations of experts solving a weapon-to-target assignment problem and a hospital resource allocation problem. We also demonstrate that policies learned from human scheduling demonstration via apprenticeship learning can substantially improve the efficiency of a branch-and-bound search for an optimal schedule. We employ this human-machine collaborative optimization technique on a variant of the weapon-to-target assignment problem. We demonstrate that this technique generates solutions substantially superior to those produced by human domain experts at a rate up to 9.5 times faster than an optimization approach and can be applied to optimally solve problems twice as complex as those solved by a human demonstrator.

In this work, we propose a new approach to interactive image search by introducing a novel form of user feedback based on natural language. The proposed approach, which we call dialog-based interactive retrieval, enables users to directly express in natural language, what the most prominent visual attributes of the image they have in mind are, improving image search results and allowing for a more natural human-computer interaction. We formulate the task as a reinforcement learning (RL) problem, and train a dialog system that takes natural language responses as user input, and produces retrieved images as output. We train this system by directly optimizing the rank of the target image, which is a non-differentiable objective. To avoid the cumbersome, inefficient, and costly process of collecting and annotating human-machine dialogs as the system learns, we utilize a model-based reinforcement learning approach by training a user simulator based on human-written relative descriptions.

It's fair to say that our world has reached a point where technology is so advanced that robots are almost expected to be lifelike – but what about robots that develop mental illnesses, hallucinations and depression like human beings do? Is this just science fiction, or can we really expect artificial intelligence to grow even more similar to humans in the not-so-distant future? Back in March, New York University hosted a symposium in New York City called Canonical Computations in Brains and Machines, where a group of neuroscientists and experts in the field of artificial intelligence spoke about overlaps in the ways in which human beings and machines think and process information. According to one of these neuroscientists – Zachary Mainen of the Champalimaud Centre for the Unknown – we might expect advanced machines to soon be able to experience some of the same mental problems that people do. "I'm drawing on the field of computational psychiatry, which assumes we can learn about a patient who's depressed or hallucinating from studying AI algorithms like reinforcement learning. If you reverse the arrow, why wouldn't an AI be subject to the sort of things that go wrong with patients?"

Classical methods to control heating systems are often marred by suboptimal performance, inability to adapt to dynamic conditions and unreasonable assumptions e.g. existence of building models. This paper presents a novel deep reinforcement learning algorithm which can control space heating in buildings in a computationally efficient manner, and benchmarks it against other known techniques. The proposed algorithm outperforms rule based control by between 5-10% in a simulation environment for a number of price signals. We conclude that, while not optimal, the proposed algorithm offers additional practical advantages such as faster computation times and increased robustness to non-stationarities in building dynamics.

Overview Deep Reinforcement Learning and GANs LiveLessons is an introduction to two of the most exciting topics in Deep Learning today.

What makes a game replayable over time? It offers new challenges over and over again. One way to do that is to include an AI opponent that is so skilled, even advanced players will continue to be challenged after hundreds of hours of play. Race has been one of the top selling boardgames this year partly because of the neural network that powers its AI. Race for the Galaxy uses a temporal difference neural network.

In reinforcement learning (RL), stochastic environments can make learning a policy difficult due to high degrees of variance. As such, variance reduction methods have been investigated in other works, such as advantage estimation and control-variates estimation. Here, we propose to learn a separate reward estimator to train the value function, to help reduce variance caused by a noisy reward signal. This results in theoretical reductions in variance in the tabular case, as well as empirical improvements in both the function approximation and tabular settings in environments where rewards are stochastic. To do so, we use a modified version of Advantage Actor Critic (A2C) on variations of Atari games.

To operate effectively in tomorrow's smart cities, autonomous vehicles (AVs) must rely on intra-vehicle sensors such as camera and radar as well as inter-vehicle communication. Such dependence on sensors and communication links exposes AVs to cyber-physical (CP) attacks by adversaries that seek to take control of the AVs by manipulating their data. Thus, to ensure safe and optimal AV dynamics control, the data processing functions at AVs must be robust to such CP attacks. To this end, in this paper, the state estimation process for monitoring AV dynamics, in presence of CP attacks, is analyzed and a novel adversarial deep reinforcement learning (RL) algorithm is proposed to maximize the robustness of AV dynamics control to CP attacks. The attacker's action and the AV's reaction to CP attacks are studied in a game-theoretic framework. In the formulated game, the attacker seeks to inject faulty data to AV sensor readings so as to manipulate the inter-vehicle optimal safe spacing and potentially increase the risk of AV accidents or reduce the vehicle flow on the roads. Meanwhile, the AV, acting as a defender, seeks to minimize the deviations of spacing so as to ensure robustness to the attacker's actions. Since the AV has no information about the attacker's action and due to the infinite possibilities for data value manipulations, the outcome of the players' past interactions are fed to long-short term memory (LSTM) blocks. Each player's LSTM block learns the expected spacing deviation resulting from its own action and feeds it to its RL algorithm. Then, the the attacker's RL algorithm chooses the action which maximizes the spacing deviation, while the AV's RL algorithm tries to find the optimal action that minimizes such deviation.

Deep Reinforcement Learning (RL) recently emerged as one of the most competitive approaches for learning in sequential decision making problems with fully observable environments, e.g., computer Go. However, very little work has been done in deep RL to handle partially observable environments. We propose a new architecture called Action-specific Deep Recurrent Q-Network (ADRQN) to enhance learning performance in partially observable domains. Actions are encoded by a fully connected layer and coupled with a convolutional observation to form an action-observation pair. The time series of action-observation pairs are then integrated by an LSTM layer that learns latent states based on which a fully connected layer computes Q-values as in conventional Deep Q-Networks (DQNs). We demonstrate the effectiveness of our new architecture in several partially observable domains, including flickering Atari games.