Temporal abstraction is key to scaling up learning and planning in reinforcement learning. While planning with temporally extended actions is well understood, creating such abstractions autonomously from data has remained challenging.We tackle this problem in the framework of options [Sutton,Precup and Singh, 1999; Precup, 2000]. We derive policy gradient theorems for options and propose a new option-critic architecture capable of learning both the internal policies and the termination conditions of options, in tandem with the policy over options, and without the need to provide any additional rewards or subgoals. Experimental results in both discrete and continuous environments showcase the flexibility and efficiency of the framework.

We propose a new method for transferring a policy from a source task to a target task in model-based reinforcement learning. Our work is motivated by scenarios where a robotic agent operates in similar but challenging environments, such as hospital wards, differentiated by structural arrangements or obstacles, such as furniture. We address problems that require fast responses adapted from incomplete, prior knowledge of the agent in new scenarios. We present an efficient selective exploration strategy that maximally reuses the source task policy. Reuse efficiency is effected through identifying sub-spaces that are different in the target environment, thus limiting the exploration needed in the target task. We empirically show that SEAPoT performs better in terms of jump starts and cumulative average rewards, as compared to existing state-of-the-art policy reuse methods.

In contrast, Inverse reinforcement learning (IRL), inverse optimal control, a discrete approach guarantees global optimality once and imitation learning(Ng and Russell 2000; Abbeel proper discrete state space is given, hence it is more suitable and Ng 2004) are modeling frameworks for acquiring rewards for driving behavior modeling. In a discrete approach, (or cost) of a certain environment by using the optimal the calculation cost of MaxEnt IRL is O( S A), where S path under a possibly different environment as training is the number of states and A is the number of actions data. In particular, in human behavior modeling, it is (Ziebart and others 2008). That is, the key for fast prediction shown that human-centered rewards can be obtained with is suppressing the increase of S depending on dimensions maximum entropy inverse reinforcement learning (MaxEnt and preparing a necessary and sufficient action set, A, IRL)(Ziebart and others 2008), which allows suboptimal for representing driving behavior. As examples of existing training data (Huang et al. 2015; Vernaza and Bagnell 2012; discretization schemes, there are mesh grid representation Dragan and Srinivasa 2012; Walker, Gupta, and Hebert (Shimosaka, Kaneko, and Nishi 2014) and random graph 2014). For instance, Ziebart et al. (Ziebart et al. 2008) modeled based representation connected with neighbors (Byravan et the driving behavior of expert taxi drivers and enabled al. 2015). In these approaches, however, A for general dynamic driving behavior prediction based on the experts' very own systems is not trivial. This is because neighbors on experience or knowledge. MaxEnt IRL based driving behavior state space defined by Euclidean distance do not necessarily prediction, which balances safety, comfort, and economic correspond to the transition area of general dynamics performance, is very promising.

This article addresses a particular Transfer Reinforcement Learning (RL) problem: when dynamics do not change from one task to another, and only the reward function does. Our method relies on two ideas, the first one is that transition samples obtained from a task can be reused to learn on any other task: an immediate reward estimator is learnt in a supervised fashion and for each sample, the reward entry is changed by its reward estimate. The second idea consists in adopting the optimism in the face of uncertainty principle and to use upper bound reward estimates. Our method is tested on a navigation task, under four Transfer RL experimental settings: with a known reward function, with strong and weak expert knowledge on the reward function, and with a completely unknown reward function. It is also evaluated in a Multi-Task RL experiment and compared with the state-of-the-art algorithms. Results reveal that this method constitutes a major improvement for transfer/multi-task problems that share dynamics.

One of the long standing goals of Artificial Intelligence (AI) is to build cognitive agents which can perform complex tasks from raw sensory inputs without explicit supervision. Recent progress in combining Reinforcement Learning objective functions and Deep Learning architectures has achieved promising results for such tasks. An important aspect of such sequential decision making problems, which has largely been neglected, is for the agent to decide on the duration of time for which to commit to actions. Such action repetition is important for computational efficiency, which is necessary for the agent to respond in real-time to events (in applications such as self-driving cars). Action Repetition arises naturally in real life as well as simulated environments. The time scale of executing an action enables an agent (both humans and AI) to decide the granularity of control during task execution. Current state of the art Deep Reinforcement Learning models, whether they are off-policy or on-policy, consist of a framework with a static action repetition paradigm, wherein the action decided by the agent is repeated for a fixed number of time steps regardless of the contextual state while executing the task. In this paper, we propose a new framework - Dynamic Action Repetition which changes Action Repetition Rate (the time scale of repeating an action) from a hyper-parameter of an algorithm to a dynamically learnable quantity. At every decision-making step, our models allow the agent to commit to an action and the time scale of executing the action. We show empirically that such a dynamic time scale mechanism improves the performance on relatively harder games in the Atari 2600 domain, independent of the underlying Deep Reinforcement Learning algorithm used.

Advances in deep reinforcement learning have allowed autonomous agents to perform well on Atari games, often outperforming humans, using only raw pixels to make their decisions. However, most of these games take place in 2D environments that are fully observable to the agent. In this paper, we present the first architecture to tackle 3D environments in first-person shooter games, that involve partially observable states. Typically, deep reinforcement learning methods only utilize visual input for training. We present a method to augment these models to exploit game feature information such as the presence of enemies or items, during the training phase. Our model is trained to simultaneously learn these features along with minimizing a Q-learning objective, which is shown to dramatically improve the training speed and performance of our agent. Our architecture is also modularized to allow different models to be independently trained for different phases of the game. We show that the proposed architecture substantially outperforms built-in AI agents of the game as well as average humans in deathmatch scenarios.

The family of temporal difference (TD) methods span a spectrum from computationally frugal linear methods like TD(λ) to data efficient least squares methods. Least square methods make the best use of available data directly computing the TD solution and thus do not require tuning a typically highly sensitive learning rate parameter, but require quadratic computation and storage. Recent algorithmic developments have yielded several sub-quadratic methods that use an approximation to the least squares TD solution, but incur bias. In this paper, we propose a new family of accelerated gradient TD (ATD) methods that (1) provide similar data efficiency benefits to least-squares methods, at a fraction of the computation and storage (2) significantly reduce parameter sensitivity compared to linear TD methods, and (3) are asymptotically unbiased. We illustrate these claims with a proof of convergence in expectation and experiments on several benchmark domains and a large-scale industrial energy allocation domain.

We propose a model-based approach to hierarchical reinforcement learning that exploits shared knowledge and selective execution at different levels of abstraction, to efficiently solve large, complex problems. Our framework adopts a new transition dynamics learning algorithm that identifies the common action-feature combinations of the subtasks, and evaluates the subtask execution choices through simulation. The framework is sample efficient, and tolerates uncertain and incomplete problem characterization of the subtasks. We test the framework on common benchmark problems and complex simulated robotic environments. It compares favorably against the state-of-the-art algorithms, and scales well in very large problems.

The stochastic shortest path problem is of crucial importance for the development of sustainable transportation systems. Existing methods based on the probability tail model seek for the path that maximizes the probability of arriving at the destination before a deadline. However, they suffer from low accuracy and/or high computational cost. We design a novel Q-learning method where the converged Q-values have the practical meaning as the actual probabilities of arriving on time so as to improve accuracy. By further adopting dynamic neural networks to learn the value function, our method can scale well to large road networks with arbitrary deadlines. Experimental results on real road networks demonstrate the significant advantages of our method over other counterparts.

When an agent cannot represent a perfectly accurate model of its environment's dynamics, model-based reinforcement learning (MBRL) can fail catastrophically. Planning involves composing the predictions of the model; when flawed predictions are composed, even minor errors can compound and render the model useless for planning. Hallucinated Replay (Talvitie 2014) trains the model to "correct" itself when it produces errors, substantially improving MBRL with flawed models. This paper theoretically analyzes this approach, illuminates settings in which it is likely to be effective or ineffective, and presents a novel error bound, showing that a model's ability to self-correct is more tightly related to MBRL performance than one-step prediction error. These results inspire an MBRL algorithm for deterministic MDPs with performance guarantees that are robust to model class limitations.