Deep reinforcement learning (DRL) methods such as the Deep Q-Network (DQN) have achieved state-of-the-art results in a variety of challenging, high-dimensional domains. This success is mainly attributed to the power of deep neural networks to learn rich domain representations for approximating the value function or policy. Batch reinforcement learning methods with linear representations, on the other hand, are more stable and require less hyper parameter tuning. Yet, substantial feature engineering is necessary to achieve good results. In this work we propose a hybrid approach -- the Least Squares Deep Q-Network (LS-DQN), which combines rich feature representations learned by a DRL algorithm with the stability of a linear least squares method. We do this by periodically re-training the last hidden layer of a DRL network with a batch least squares update. Key to our approach is a Bayesian regularization term for the least squares update, which prevents over-fitting to the more recent data. We tested LS-DQN on five Atari games and demonstrate significant improvement over vanilla DQN and Double-DQN. We also investigated the reasons for the superior performance of our method. Interestingly, we found that the performance improvement can be attributed to the large batch size used by the LS method when optimizing the last layer.

Monotonic policy improvement and off-policy learning are two main desirable properties for reinforcement learning algorithms. In this paper, by lower bounding the performance difference of two policies, we show that the monotonic policy improvement is guaranteed from on- and off-policy mixture samples. An optimization procedure which applies the proposed bound can be regarded as an off-policy natural policy gradient method. In order to support the theoretical result, we provide a trust region policy optimization method using experience replay as a naive application of our bound, and evaluate its performance in two classical benchmark problems.

There has recently been significant interest in hard attention models for tasks such as object recognition, visual captioning and speech recognition. Hard attention can offer benefits over soft attention such as decreased computational cost, but training hard attention models can be difficult because of the discrete latent variables they introduce. Previous work used REINFORCE and Q-learning to approach these issues, but those methods can provide high-variance gradient estimates and be slow to train. In this paper, we tackle the problem of learning hard attention for a sequential task using variational inference methods, specifically the recently introduced VIMCO and NVIL. Furthermore, we propose a novel baseline that adapts VIMCO to this setting. We demonstrate our method on a phoneme recognition task in clean and noisy environments and show that our method outperforms REINFORCE, with the difference being greater for a more complicated task.

Robots interacting with humans often have to represent and reason with different descriptions of incomplete domain knowledge and uncertainty, and revise this knowledge over time. Towards achieving these capabilities, the architecture described in this paper combines the complementary strengths of declarative programming, probabilistic graphical models, and reinforcement learning. For any given goal, non-monotonic logical reasoning with a coarse-resolution representation of the domain is used to compute a tentative plan of abstract actions. Each abstract action is implemented as a sequence of concrete actions by reasoning probabilistically over the relevant part of a fine-resolution representation tightly-coupled to the coarse-resolution representation. The outcomes of executing the concrete actions are used for subsequenct reasoning at the coarse resolution. Furthermore, the task of interactively learning axioms governing action capabilities, preconditions and effects, is posed as a relational reinforcement learning problem, using decision tree regression and sampling to construct and generalize over candidate axioms. These capabilities are illustrated in simulation and on a physical robot moving objects to specific people or locations in an indoor domain.

This paper proposes a reinterpretation of selective attention as a form of control of working memory based on self-generated reward signals and model-free reinforcement learning. In addition to being simple and parsimonious, this approach systematizes a number of classic psychological constructs without calling for additional, specific mechanisms. Finally, the papers presents the results of an empirical test of this framework, and elaborates on the implications of our findings for general models of control and intelligent behavior, as well as neurobiological models of the basal ganglia.

In recent artificial intelligence (AI) research, convolutional neural networks (CNNs) can create artificial agents capable of self-learning. Self-learning autonomous moving objects utilize machine vision techniques based on processing and recognizing objects in digital images. Afterwards, deep reinforcement learning (Deep-RL) is applied to understand and learn intelligent actions and controls. The objective of our research is to study methods and designs on how machine vision and deep machine learning algorithms can be implemented in a virtual world (e.g., a computer game) for moving objects (e.g., vehicles or aircrafts) to improve their navigation and detection of threats in real life. In this paper, we create a framework for generating and using data from computer games to be used in CNNs and Deep-RL to perform intelligent actions. We show the initial results of applying the framework and identify various military applications that may benefit from this research.

In this position paper, we propose a game theoretic formulation of the adversarial learning problem called a RepeatedBayesian Stackelberg Game (RBSG) that can be used by aprediction mechanism to make itself robust against adversarial examples.

Social interacting is a complex task for which machine learning holds particular promise. However, as no sufficiently accurate simulator of human interactions exists today, the learning of social interaction strategies has to happen online in the real world. Actions executed by the robot impact on humans, and as such have to be carefully selected, making it impossible to rely on random exploration. Additionally, no clear reward function exists for social interactions. This implies that traditional approaches used for Reinforcement Learning cannot be directly applied for learning how to interact with the social world. As such we argue that robots will profit from human expertise and guidance to learn social interactions. However, as the quantity of input a human can provide is limited, new methods have to be designed to use human input more efficiently. In this paper we describe a setup in which we combine a framework called Supervised Progressively Autonomous Robot Competencies (SPARC), which allows safer online learning with Reinforcement Learning, with the use of partial states rather than full states to accelerate generalisation and obtain a usable action policy more quickly.

Learning from demonstration is a popular method for teaching robots new skills. However, little work has looked at how to measure safety in the context of learning from demonstrations. We discuss three different types of safety problems that are important for robot learning from human demonstrations: (1) using demonstrations to evaluate the safety of a robot's current policy, (2) using demonstrations to enable risk-aware policy improvement, and (3) determining when the demonstrations received by the robot are sufficient to ensure a desired safety level. We propose a risk-aware Bayesian sampling approach based on inverse reinforcement learning that provides a first step towards addressing these problems. We demonstrate the validity of our approach on a simulated navigation task and discuss promising areas for future work.

Despite increasing attention paid to the need for fast, scalable methods to analyze next-generation neuroscience data, comparatively little attention has been paid to the development of similar methods for behavioral analysis. Just as the volume and complexity of brain data have grown, behavioral paradigms in systems neuroscience have likewise become more naturalistic and less constrained, necessitating an increase in the flexibility and scalability of the models used to study them. In particular, key assumptions made in the analysis of typical decision paradigms --- optimality; analytic tractability; discrete, low-dimensional action spaces --- may be untenable in richer tasks. Here, using the case of a two-player, real-time, continuous strategic game as an example, we show how the use of modern machine learning methods allows us to relax each of these assumptions. Following an inverse reinforcement learning approach, we are able to succinctly characterize the joint distribution over players' actions via a generative model that allows us to simulate realistic game play. We compare simulated play from a number of generative time series models and show that ours successfully resists mode collapse while generating trajectories with the rich variability of real behavior. Together, these methods offer a rich class of models for the analysis of continuous action tasks at the single-trial level.