This paper presents the first two editions of Visual Doom AI Competition, held in 2016 and 2017. The challenge was to create bots that compete in a multi-player deathmatch in a first-person shooter (FPS) game, Doom. The bots had to make their decisions based solely on visual information, i.e., a raw screen buffer. To play well, the bots needed to understand their surroundings, navigate, explore, and handle the opponents at the same time. These aspects, together with the competitive multi-agent aspect of the game, make the competition a unique platform for evaluating the state of the art reinforcement learning algorithms. The paper discusses the rules, solutions, results, and statistics that give insight into the agents' behaviors. Best-performing agents are described in more detail. The results of the competition lead to the conclusion that, although reinforcement learning can produce capable Doom bots, they still are not yet able to successfully compete against humans in this game. The paper also revisits the ViZDoom environment, which is a flexible, easy to use, and efficient 3D platform for research for vision-based reinforcement learning, based on a well-recognized first-person perspective game Doom.

We propose an expert-augmented actor-critic algorithm, which we evaluate on two environments with sparse rewards: Montezumas Revenge and a demanding maze from the ViZDoom suite. In the case of Montezumas Revenge, an agent trained with our method achieves very good results consistently scoring above 27,000 points (in many experiments beating the first world). With an appropriate choice of hyperparameters, our algorithm surpasses the performance of the expert data. In a number of experiments, we have observed an unreported bug in Montezumas Revenge which allowed the agent to score more than 800,000 points.

Deep reinforcement learning can match and exceed human performance, but if even minor changes are introduced to the environment artificial networks often can't adapt. Humans meanwhile are quite adaptable. We hypothesize that this is partly because of how humans use heuristics, and partly because humans can imagine new and more challenging environments to learn from. We've developed a model of hierarchical reinforcement learning that combines both these elements into a stumbler-strategist network. We test transfer performance of this network using Wythoff's game, a gridworld environment with a known optimal strategy. We show that combining imagined play with a heuristic--labeling each position as "good" or "bad"'--both accelerates learning and promotes transfer to novel games, while also improving model interpretability.

Autonomous vehicles should be able to generate accurate probabilistic predictions for uncertain behavior of other road users. Moreover, reactive predictions are necessary in highly interactive driving scenarios to answer "what if I take this action in the future" for autonomous vehicles. There is no existing unified framework to homogenize the problem formulation, representation simplification, and evaluation metric for various prediction methods, such as probabilistic graphical models (PGM), neural networks (NN) and inverse reinforcement learning (IRL). In this paper, we formulate a probabilistic reaction prediction problem, and reveal the relationship between reaction and situation prediction problems. We employ prototype trajectories with designated motion patterns other than "intention" to homogenize the representation so that probabilities corresponding to each trajectory generated by different methods can be evaluated. We also discuss the reasons why "intention" is not suitable to serve as a motion indicator in highly interactive scenarios. We propose to use Brier score as the baseline metric for evaluation. In order to reveal the fatality of the consequences when the predictions are adopted by decision-making and planning, we propose a fatality-aware metric, which is a weighted Brier score based on the criticality of the trajectory pairs of the interacting entities. Conservatism and non-defensiveness are defined from the weighted Brier score to indicate the consequences caused by inaccurate predictions. Modified methods based on PGM, NN and IRL are provided to generate probabilistic reaction predictions in an exemplar scenario of nudging from a highway ramp. The results are evaluated by the baseline and proposed metrics to construct a mini benchmark. Analysis on the properties of each method is also provided by comparing the baseline and proposed metric scores.

Finding tight bounds on the optimal solution is a critical element of practical solution methods for discrete optimization problems. In the last decade, decision diagrams (DDs) have brought a new perspective on obtaining upper and lower bounds that can be significantly better than classical bounding mechanisms, such as linear relaxations. It is well known that the quality of the bound achieved through this flexible bounding method is highly reliant on the ordering of variables chosen for building the diagram, and finding an ordering that optimizes standard metrics, or even improving one, is an NP-hard problem. In this paper, we propose an innovative and generic approach based on deep reinforcement learning for obtaining an ordering for tightening the bounds obtained with relaxed and restricted DDs. We apply the approach to both the Maximum Independent Set Problem and the Maximum Cut Problem. Experimental results on synthetic instances show that the deep reinforcement learning approach, by achieving tighter objective function bounds, generally outperforms ordering methods commonly used in the literature when the distribution of instances is known. To the best knowledge of the authors, this is the first paper to apply machine learning to directly improve relaxation bounds obtained by general-purpose bounding mechanisms for combinatorial optimization problems.

In online display advertising, guaranteed contracts and real-time bidding (RTB) are two major ways to sell impressions for a publisher. Despite the increasing popularity of RTB, there is still half of online display advertising revenue generated from guaranteed contracts. Therefore, simultaneously selling impressions through both guaranteed contracts and RTB is a straightforward choice for a publisher to maximize its yield. However, deriving the optimal strategy to allocate impressions is not a trivial task, especially when the environment is unstable in real-world applications. In this paper, we formulate the impression allocation problem as an auction problem where each contract can submit virtual bids for individual impressions. With this formulation, we derive the optimal impression allocation strategy by solving the optimal bidding functions for contracts. Since the bids from contracts are decided by the publisher, we propose a multi-agent reinforcement learning (MARL) approach to derive cooperative policies for the publisher to maximize its yield in an unstable environment. The proposed approach also resolves the common challenges in MARL such as input dimension explosion, reward credit assignment, and non-stationary environment. Experimental evaluations on large-scale real datasets demonstrate the effectiveness of our approach.

This paper proposes a low-cost, easily realizable strategy to equip a reinforcement learning (RL) agent the capability of behaving ethically. Our model allows the designers of RL agents to solely focus on the task to achieve, without having to worry about the implementation of multiple trivial ethical patterns to follow. Based on the assumption that the majority of human behavior, regardless which goals they are achieving, is ethical, our design integrates human policy with the RL policy to achieve the target objective with less chance of violating the ethical code that human beings normally obey.

While deep reinforcement learning (deep RL) agents are effective at maximizing rewards, it is often unclear what strategies they use to do so. In this paper, we take a step toward explaining deep RL agents through a case study using Atari 2600 environments. In particular, we focus on using saliency maps to understand how an agent learns and executes a policy. We introduce a method for generating useful saliency maps and use it to show 1) what strong agents attend to, 2) whether agents are making decisions for the right or wrong reasons, and 3) how agents evolve during learning. We also test our method on non-expert human subjects and find that it improves their ability to reason about these agents. Overall, our results show that saliency information can provide significant insight into an RL agent's decisions and learning behavior.

Reward design, the problem of selecting an appropriate reward function for an AI system, is both critically important, as it encodes the task the system should perform, and challenging, as it requires reasoning about and understanding the agent's environment in detail. AI practitioners often iterate on the reward function for their systems in a trial-and-error process to get their desired behavior. Inverse reward design (IRD) is a preference inference method that infers a true reward function from an observed, possibly misspecified, proxy reward function. This allows the system to determine when it should trust its observed reward function and respond appropriately. This has been shown to avoid problems in reward design such as negative side-effects (omitting a seemingly irrelevant but important aspect of the task) and reward hacking (learning to exploit unanticipated loopholes). In this paper, we actively select the $\textit{set of proxy reward functions}$ available to the designer. This improves the quality of inference and simplifies the associated reward design problem. We present two types of queries: discrete queries, where the system designer chooses from a discrete set of reward functions, and feature queries, where the system queries the designer for weights on a small set of features. We evaluate this approach with experiments in a personal shopping assistant domain and a 2D navigation domain. We find that our approach leads to reduced regret at test time compared with vanilla IRD. Our results indicate that actively selecting the set of available reward functions is a promising direction to improve the efficiency and effectiveness of reward design.

Autonomous vehicles (AVs) are on the road. To safely and efficiently interact with other road participants, AVs have to accurately predict the behavior of surrounding vehicles and plan accordingly. Such prediction should be probabilistic, to address the uncertainties in human behavior. Such prediction should also be interactive, since the distribution over all possible trajectories of the predicted vehicle depends not only on historical information, but also on future plans of other vehicles that interact with it. To achieve such interaction-aware predictions, we propose a probabilistic prediction approach based on hierarchical inverse reinforcement learning (IRL). First, we explicitly consider the hierarchical trajectory-generation process of human drivers involving both discrete and continuous driving decisions. Based on this, the distribution over all future trajectories of the predicted vehicle is formulated as a mixture of distributions partitioned by the discrete decisions. Then we apply IRL hierarchically to learn the distributions from real human demonstrations. A case study for the ramp-merging driving scenario is provided. The quantitative results show that the proposed approach can accurately predict both the discrete driving decisions such as yield or pass as well as the continuous trajectories.