The never-ending saga of machines outperforming humans has a new chapter. An AI algorithm has again beaten a human fighter pilot in a virtual dogfight. The contest was the finale of the U.S. military's AlphaDogfight challenge, an effort to "demonstrate the feasibility of developing effective, intelligent autonomous agents capable of defeating adversary aircraft in a dogfight. Last August, Defense Advanced Research Project Agency, or DARPA, selected eight teams ranging from large, traditional defense contractors like Lockheed Martin to small groups like Heron Systems to compete in a series of trials in November and January. In the final, on Thursday, Heron Systems emerged as the victor against the seven other teams after two days of old school dogfights, going after each other using nose-aimed guns only. Heron then faced off against a human fighter pilot sitting in a simulator and wearing a virtual reality helmet, and won five rounds to zero. The other winner in Thursday's event was deep reinforcement learning, wherein artificial intelligence algorithms get to try out a task in a virtual environment over and over again, sometimes very quickly, until they develop something like understanding. Deep reinforcement played a key role in Heron System's agent, as well as Lockheed Martin's, the second runner up. Matt Tarascio, vice president of artificial intelligence, and Lee Ritholtz, director and chief architect of artificial intelligence, from Lockheed Martin told Defense One that trying to get an algorithm to perform well in air combat is very different than teaching software simply "to fly," or maintain a particular direction, altitude, and speed. Software will begin with a complete lack of understanding about even very basic flight tasks, explained Ritholtz, putting it at a disadvantage against any human, at first. "You don't have to teach a human [that] it shouldn't crash into the ground… They have basic instincts that the algorithm doesn't have," in terms of training. "That means dying a lot.

Episodic control provides a highly sample-efficient method for reinforcement learning while enforcing high memory and computational requirements. This work proposes a simple heuristic for reducing these requirements, and an application to Model-Free Episodic Control (MFEC) is presented. Experiments on Atari games show that this heuristic successfully reduces MFEC computational demands while producing no significant loss of performance when conservative choices of hyperparameters are used. Consequently, episodic control becomes a more feasible option when dealing with reinforcement learning tasks.

Humans generally use natural language to communicate task requirements amongst each other. It is desirable that this would be similar for autonomous machines (e.g. robots) such that humans can convey goals or assign tasks more easily. However, understanding natural language goals and mapping them to sequences of states and actions is challenging. Previous research has encountered difficulty generalizing learned policies to new natural language goals and environments. In this paper, we propose an adversarial inverse reinforcement learning algorithm that learns a language-conditioned policy and reward function. To improve the generalization of the learned policy and reward function, we use a variational goal generator that relabels trajectories and samples diverse goals during training. Our algorithm outperforms baselines by a large margin on a vision-based natural language instruction following dataset, demonstrating a promising advance in providing natural language instructions to agents without reliance on instruction templates.

It is not until we become senior citizens do we recognise how much we took maintaining a simple standing posture for granted. It is truly fascinating to observe the magnitude of control the human brain exercises, in real time, to activate and deactivate the lower body muscles and solve a multi-link 3D inverted pendulum problem in order to maintain a stable standing posture. This realisation is even more apparent when training an artificial intelligence (AI) agent to maintain a standing posture of a digital musculoskeletal avatar due to the error propagation problem. In this work we address the error propagation problem by introducing an iterative training procedure for deep reinforcement learning which allows the agent to learn a finite set of actions and how to coordinate between them in order to achieve a stable standing posture. The proposed training approach allowed the agent to increase standing duration from 4 seconds using the traditional training method to 348 seconds using the proposed method. The proposed training method allowed the agent to generalise and accommodate perception and actuation noise for almost 108 seconds.

Learning complex tasks from scratch is challenging and often impossible for humans as well as for artificial agents. A curriculum can be used instead, which decomposes a complex task (target task) into a sequence of source tasks (the curriculum). Each source task is a simplified version of the next source task with increasing complexity. Learning then occurs gradually by training on each source task while using knowledge from the curriculum's prior source tasks. In this study, we present a new algorithm that combines curriculum learning with Hindsight Experience Replay (HER), to learn sequential object manipulation tasks for multiple goals and sparse feedback. The algorithm exploits the recurrent structure inherent in many object manipulation tasks and implements the entire learning process in the original simulation without adjusting it to each source task. We have tested our algorithm on three challenging throwing tasks and show vast improvements compared to vanilla-HER.

Ensuring quality of service (QoS) guarantees in service systems is a challenging task, particularly when the system is composed of more fine-grained services, such as service function chains. An important QoS metric in service systems is the end-to-end delay, which becomes even more important in delay-sensitive applications, where the jobs must be completed within a time deadline. Admission control is one way of providing end-to-end delay guarantee, where the controller accepts a job only if it has a high probability of meeting the deadline. In this paper, we propose a reinforcement learning-based admission controller that guarantees a probabilistic upper-bound on the end-to-end delay of the service system, while minimizes the probability of unnecessary rejections. Our controller only uses the queue length information of the network and requires no knowledge about the network topology or system parameters. Since long-term performance metrics are of great importance in service systems, we take an average-reward reinforcement learning approach, which is well suited to infinite horizon problems. Our evaluations verify that the proposed RL-based admission controller is capable of providing probabilistic bounds on the end-to-end delay of the network, without using system model information.

Developing agents that can perform challenging complex tasks is the goal of reinforcement learning. The model-free reinforcement learning has been considered as a feasible solution. However, the state of the art research has been to develop increasingly complicated techniques. This increasing complexity makes the reconstruction difficult. Furthermore, the problem of reward dependency is still exists. As a result, research on imitation learning, which learns policy from a demonstration of experts, has begun to attract attention. Imitation learning directly learns policy based on data on the behavior of the experts without the explicit reward signal provided by the environment. However, imitation learning tries to optimize policies based on deep reinforcement learning such as trust region policy optimization. As a result, deep reinforcement learning based imitation learning also poses a crisis of reproducibility. The issue of complex model-free model has received considerable critical attention. A derivative-free optimization based reinforcement learning and the simplification on policies obtain competitive performance on the dynamic complex tasks. The simplified policies and derivative free methods make algorithm be simple. The reconfiguration of research demo becomes easy. In this paper, we propose an imitation learning method that takes advantage of the derivative-free optimization with simple linear policies. The proposed method performs simple random search in the parameter space of policies and shows computational efficiency. Experiments in this paper show that the proposed model, without a direct reward signal from the environment, obtains competitive performance on the MuJoCo locomotion tasks.

In audio-visual navigation, an agent intelligently travels through a complex, unmapped 3D environment using both sights and sounds to find a sound source (e.g., a phone ringing in another room). Existing models learn to act at a fixed granularity of agent motion and rely on simple recurrent aggregations of the audio observations. We introduce a reinforcement learning approach to audio-visual navigation with two key novel elements 1) audio-visual waypoints that are dynamically set and learned end-to-end within the navigation policy, and 2) an acoustic memory that provides a structured, spatially grounded record of what the agent has heard as it moves. Both new ideas capitalize on the synergy of audio and visual data for revealing the geometry of an unmapped space. We demonstrate our approach on the challenging Replica environments of real-world 3D scenes. Our model improves the state of the art by a substantial margin, and our experiments reveal that learning the links between sights, sounds, and space is essential for audio-visual navigation.

The general formulation of reinforcement learning can be defined by an agent interacting with an environment attempting to maximize its reward function. This is often formulated as a Markov Decision Process (MDP).

In this paper, we introduce the group concept into multi-agent reinforcement learning. In this method, agents are divided into several groups and each group completes a specific subtask so that agents can cooperate to complete the main task. Existing methods use the communication vector to exchange information between agents. This may encounter communication redundancy. To solve this problem, we propose a MARL method based on graph clustering. It allows agents to adaptively learn group features and replaces the communication operation. In our method, agent features are divide into two types, including in-group features and individual features. They represent the generality and differences between agents, respectively. Based on the graph attention network(GAT), we introduce the graph clustering method as a punishment to optimize agent group feature. Then these features are used to generate individual Q value. To overcome the consistent problem brought by GAT, we introduce the split loss to distinguish agent features. Our method is easy to convert into the CTDE framework via using Kullback-Leibler divergence method. Empirical results are evaluated on a challenging set of StarCraft II micromanagement tasks. The result shows that our method outperforms existing multi-agent reinforcement learning methods and the performance increases with the number of agents increasing.