Integrating simulation models and deep reinforcement learning, this position … Knowledge of quantum technology or machine learning related to the …

Recent work in deep reinforcement learning (RL) has produced algorithms capable of mastering challenging games such as Go, chess, or shogi. In these works the RL agent directly observes the natural state of the game and controls that state directly with its actions. However, when humans play such games, they do not just reason about the moves but also interact with their physical environment. They understand the state of the game by looking at the physical board in front of them and modify it by manipulating pieces using touch and fine-grained motor control. Mastering complicated physical systems with abstract goals is a central challenge for artificial intelligence, but it remains out of reach for existing RL algorithms.

A remarkable characteristic of human intelligence is our ability to learn tasks quickly. Most humans can learn reasonably complex skills like tool-use and gameplay within just a few hours, and understand the basics after only a few attempts. This suggests that data-efficient learning may be a meaningful part of developing broader intelligence. On the other hand, Deep Reinforcement Learning (RL) algorithms can achieve superhuman performance on games like Atari, Starcraft, Dota, and Go, but require large amounts of data to get there. Achieving superhuman performance on Dota took over 10,000 human years of gameplay. Unlike simulation, skill acquisition in the real-world is constrained to wall-clock time.

In an effort to overcome limitations of reward-driven feature learning in deep reinforcement learning (RL) from images, we propose decoupling representation learning from policy learning. To this end, we introduce a new unsupervised learning (UL) task, called Augmented Temporal Contrast (ATC), which trains a convolutional encoder to associate pairs of observations separated by a short time difference, under image augmentations and using a contrastive loss. In online RL experiments, we show that training the encoder exclusively using ATC matches or outperforms end-to-end RL in most environments. Additionally, we benchmark several leading UL algorithms by pre-training encoders on expert demonstrations and using them, with weights frozen, in RL agents; we find that agents using ATC-trained encoders outperform all others. We also train multi-task encoders on data from multiple environments and show generalization to different downstream RL tasks. Finally, we ablate components of ATC, and introduce a new data augmentation to enable replay of (compressed) latent images from pre-trained encoders when RL requires augmentation. Ever since the first fully-learned approach succeeded at playing Atari games from screen images (Mnih et al., 2015), standard practice in deep reinforcement learning (RL) has been to learn visual features and a control policy jointly, end-to-end.

We address a critical yet largely unsolved anomaly detection problem, in which we aim to learn detection models from a small set of partially labeled anomalies and a large-scale unlabeled dataset. This is a common scenario in many important applications. Existing related methods either proceed unsupervised with the unlabeled data, or exclusively fit the limited anomaly examples that often do not span the entire set of anomalies. We propose here instead a deep reinforcement-learning-based approach that actively seeks novel classes of anomaly that lie beyond the scope of the labeled training data. This approach learns to balance exploiting its existing data model against exploring for new classes of anomaly. It is thus able to exploit the labeled anomaly data to improve detection accuracy, without limiting the set of anomalies sought to those given anomaly examples. This is of significant practical benefit, as anomalies are inevitably unpredictable in form and often expensive to miss. Extensive experiments on 48 real-world datasets show that our approach significantly outperforms five state-of-the-art competing methods.

We study worst-case guarantees on the expected return of fixed-dataset policy optimization algorithms. Our core contribution is a unified conceptual and mathematical framework for the study of algorithms in this regime. This analysis reveals that for naive approaches, the possibility of erroneous value overestimation leads to a difficult-to-satisfy requirement: in order to guarantee that we select a policy which is near-optimal, we may need the dataset to be informative of the value of every policy. To avoid this, algorithms can follow the pessimism principle, which states that we should choose the policy which acts optimally in the worst possible world. We show why pessimistic algorithms can achieve good performance even when the dataset is not informative of every policy, and derive families of algorithms which follow this principle. These theoretical findings are validated by experiments on a tabular gridworld, and deep learning experiments on four MinAtar environments.

Artificial behavioral agents are often evaluated based on their consistent behaviors and performance to take sequential actions in an environment to maximize some notion of cumulative reward. However, human decision making in real life usually involves different strategies and behavioral trajectories that lead to the same empirical outcome. Motivated by clinical literature of a wide range of neurological and psychiatric disorders, we propose here a more general and flexible parametric framework for sequential decision making that involves a two-stream reward processing mechanism. We demonstrated that this framework is flexible and unified enough to incorporate a family of problems spanning multi-armed bandits (MAB), contextual bandits (CB) and reinforcement learning (RL), which decompose the sequential decision making process in different levels. Inspired by the known reward processing abnormalities of many mental disorders, our clinically-inspired agents demonstrated interesting behavioral trajectories and comparable performance on simulated tasks with particular reward distributions, a real-world dataset capturing human decision-making in gambling tasks, and the PacMan game across different reward stationarities in a lifelong learning setting.

Deep reinforcement learning offers a model-free alternative to supervised deep learning and classical optimization for solving the transmit power control problem in wireless networks. The multi-agent deep reinforcement learning approach considers each transmitter as an individual learning agent that determines its transmit power level by observing the local wireless environment. Following a certain policy, these agents learn to collaboratively maximize a global objective, e.g., a sum-rate utility function. This multi-agent scheme is easily scalable and practically applicable to large-scale cellular networks. In this work, we present a distributively executed continuous power control algorithm with the help of deep actor-critic learning, and more specifically, by adapting deep deterministic policy gradient. Furthermore, we integrate the proposed power control algorithm to a time-slotted system where devices are mobile and channel conditions change rapidly. We demonstrate the functionality of the proposed algorithm using simulation results.

Off-policy policy optimization is a challenging problem in reinforcement learning (RL). The algorithms designed for this problem often suffer from high variance in their estimators, which results in poor sample efficiency, and have issues with convergence. A few variance-reduced on-policy policy gradient algorithms have been recently proposed that use methods from stochastic optimization to reduce the variance of the gradient estimate in the REINFORCE algorithm. However, these algorithms are not designed for the off-policy setting and are memory-inefficient, since they need to collect and store a large ``reference'' batch of samples from time to time. To achieve variance-reduced off-policy-stable policy optimization, we propose an algorithm family that is memory-efficient, stochastically variance-reduced, and capable of learning from off-policy samples. Empirical studies validate the effectiveness of the proposed approaches.

A COVID-19 vaccine is our best bet for mitigating the ongoing onslaught of the pandemic. However, vaccine is also expected to be a limited resource. An optimal allocation strategy, especially in countries with access inequities and a temporal separation of hot-spots might be an effective way of halting the disease spread. We approach this problem by proposing a novel pipeline VacSIM that dovetails Actor-Critic using Kronecker-Factored Trust Region (ACKTR) model into a Contextual Bandits approach for optimizing the distribution of COVID-19 vaccine. Whereas the ACKTR model suggests better actions and rewards, Contextual Bandits allow online modifications that may need to be implemented on a day-to-day basis in the real world scenario. We evaluate this framework against a naive allocation approach of distributing vaccine proportional to the incidence of COVID-19 cases in five different States across India and demonstrate up to 100,000 additional lives potentially saved and a five-fold increase in the efficacy of limiting the spread over a period of 30 days through the VacSIM approach. We also propose novel evaluation strategies including a standard compartmental model based projections and a causality preserving evaluation of our model. Finally, we contribute a new Open-AI environment meant for the vaccine distribution scenario, and open-source VacSIM for wide testing and applications across the globe.