However, these AI algorithms cannot explain the thought processes behind their decisions. A computer that masters protein folding and also tells researchers more about the rules of biology is much more useful than a computer that folds proteins without explanation. Therefore, AI researchers like me are now turning our efforts toward developing AI algorithms that can explain themselves in a manner that humans can understand. If we can do this, I believe that AI will be able to uncover and teach people new facts about the world that have not yet been discovered, leading to new innovations. One field of AI, called reinforcement learning, studies how computers can learn from their own experiences.

We propose a reinforcement learning (RL) approach to model optimal exercise strategies for option-type products. We pursue the RL avenue in order to learn the optimal action-value function of the underlying stopping problem. In addition to retrieving the optimal Q-function at any time step, one can also price the contract at inception. We first discuss the standard setting with one exercise right, and later extend this framework to the case of multiple stopping opportunities in the presence of constraints. We propose to approximate the Q-function with a deep neural network, which does not require the specification of basis functions as in the least-squares Monte Carlo framework and is scalable to higher dimensions. We derive a lower bound on the option price obtained from the trained neural network and an upper bound from the dual formulation of the stopping problem, which can also be expressed in terms of the Q-function. Our methodology is illustrated with examples covering the pricing of swing options.

Autoregressive sequence Generation models have achieved state-of-the-art performance in areas like machine translation and image captioning. These models are autoregressive in that they generate each word by conditioning on previously generated words, which leads to heavy latency during inference. Recently, non-autoregressive decoding has been proposed in machine translation to speed up the inference time by generating all words in parallel. Typically, these models use the word-level cross-entropy loss to optimize each word independently. However, such a learning process fails to consider the sentence-level consistency, thus resulting in inferior generation quality of these non-autoregressive models. In this paper, we propose a simple and efficient model for Non-Autoregressive sequence Generation (NAG) with a novel training paradigm: Counterfactuals-critical Multi-Agent Learning (CMAL). CMAL formulates NAG as a multi-agent reinforcement learning system where element positions in the target sequence are viewed as agents that learn to cooperatively maximize a sentence-level reward. On MSCOCO image captioning benchmark, our NAG method achieves a performance comparable to state-of-the-art autoregressive models, while brings 13.9x decoding speedup. On WMT14 EN-DE machine translation dataset, our method outperforms cross-entropy trained baseline by 6.0 BLEU points while achieves the greatest decoding speedup of 17.46x.

With the decreasing cost of data collection, the space of variables or features that can be used to characterize a particular predictor of interest continues to grow exponentially. Therefore, identifying the most characterizing features that minimizes the variance without jeopardizing the bias of our models is critical to successfully training a machine learning model. In addition, identifying such features is critical for interpretability, prediction accuracy and optimal computation cost. While statistical methods such as subset selection, shrinkage, dimensionality reduction have been applied in selecting the best set of features, some other approaches in literature have approached feature selection task as a search problem where each state in the search space is a possible feature subset. In this paper, we solved the feature selection problem using Reinforcement Learning. Formulating the state space as a Markov Decision Process (MDP), we used Temporal Difference (TD) algorithm to select the best subset of features. Each state was evaluated using a robust and low cost classifier algorithm which could handle any non-linearities in the dataset.

We address a state-of-the-art reinforcement learning (RL) control approach to automatically configure robotic prosthesis impedance parameters to enable end-to-end, continuous locomotion intended for transfemoral amputee subjects. Specifically, our actor-critic based RL provides tracking control of a robotic knee prosthesis to mimic the intact knee profile. This is a significant advance from our previous RL based automatic tuning of prosthesis control parameters which have centered on regulation control with a designer prescribed robotic knee profile as the target. In addition to presenting the complete tracking control algorithm based on direct heuristic dynamic programming (dHDP), we provide an analytical framework for the tracking controller with constrained inputs. We show that our proposed tracking control possesses several important properties, such as weight convergence of the learning networks, Bellman (sub)optimality of the cost-to-go value function and control input, and practical stability of the human-robot system under input constraint. We further provide a systematic simulation of the proposed tracking control using a realistic human-robot system simulator, the OpenSim, to emulate how the dHDP enables level ground walking, walking on different terrains and at different paces. These results show that our proposed dHDP based tracking control is not only theoretically suitable, but also practically useful.

Dynamic platforms that operate over manyunique terrain conditions typically require multiple controllers.To transition safely between controllers, there must be anoverlap of states between adjacent controllers. We developa novel method for training Setup Policies that bridge thetrajectories between pre-trained Deep Reinforcement Learning(DRL) policies. We demonstrate our method with a simulatedbiped traversing a difficult jump terrain, where a single policyfails to learn the task, and switching between pre-trainedpolicies without Setup Policies also fails. We perform anablation of key components of our system, and show thatour method outperforms others that learn transition policies.We demonstrate our method with several difficult and diverseterrain types, and show that we can use Setup Policies as partof a modular control suite to successfully traverse a sequence ofcomplex terrains. We show that using Setup Policies improvesthe success rate for traversing a single difficult jump terrain(from 1.5%success rate without Setup Policies to 82%), and asequence of various terrains (from 6.5%without Setup Policiesto 29.1%).

The partially observable card game Hanabi has recently been proposed as a new AI challenge problem due to its dependence on implicit communication conventions and apparent necessity of theory of mind reasoning for efficient play. In this work, we propose a mechanism for imbuing Reinforcement Learning agents with a theory of mind to discover efficient cooperative strategies in Hanabi. The primary contributions of this work are threefold: First, a formal definition of a computationally tractable mechanism for computing hand probabilities in Hanabi. Second, an extension to conventional Deep Reinforcement Learning that introduces reasoning over finitely nested theory of mind belief hierarchies. Finally, an intrinsic reward mechanism enabled by theory of mind that incentivizes agents to share strategically relevant private knowledge with their teammates. We demonstrate the utility of our algorithm against Rainbow, a state-of-the-art Reinforcement Learning agent.

Deep reinforcement learning (DRL) is applied in safety-critical domains such as robotics and autonomous driving. It achieves superhuman abilities in many tasks, however whether DRL agents can be shown to act safely is an open problem. Atari games are a simple yet challenging exemplar for evaluating the safety of DRL agents and feature a diverse portfolio of game mechanics. The safety of neural agents has been studied before using methods that either require a model of the system dynamics or an abstraction; unfortunately, these are unsuitable to Atari games because their low-level dynamics are complex and hidden inside their emulator. We present the first exact method for analysing and ensuring the safety of DRL agents for Atari games. Our method only requires access to the emulator. First, we give a set of 43 properties that characterise "safe behaviour" for 30 games. Second, we develop a method for exploring all traces induced by an agent and a game and consider a variety of sources of game non-determinism. We observe that the best available DRL agents reliably satisfy only very few properties; several critical properties are violated by all agents. Finally, we propose a countermeasure that combines a bounded explicit-state exploration with shielding. We demonstrate that our method improves the safety of all agents over multiple properties.

Network slicing manages network resources as virtual resource blocks (RBs) for the 5G Radio Access Network (RAN). Each communication request comes with quality of experience (QoE) requirements such as throughput and latency/deadline, which can be met by assigning RBs, communication power, and processing power to the request. For a completed request, the achieved reward is measured by the weight (priority) of this request. Then, the reward is maximized over time by allocating resources, e.g., with reinforcement learning (RL). In this paper, we introduce a novel flooding attack on 5G network slicing, where an adversary generates fake network slicing requests to consume the 5G RAN resources that would be otherwise available to real requests. The adversary observes the spectrum and builds a surrogate model on the network slicing algorithm through RL that decides on how to craft fake requests to minimize the reward of real requests over time. We show that the portion of the reward achieved by real requests may be much less than the reward that would be achieved when there was no attack. We also show that this flooding attack is more effective than other benchmark attacks such as random fake requests and fake requests with the minimum resource requirement (lowest QoE requirement). Fake requests may be detected due to their fixed weight. As an attack enhancement, we present schemes to randomize weights of fake requests and show that it is still possible to reduce the reward of real requests while maintaining the balance on weight distributions.

We study the robustness of reinforcement learning (RL) with adversarially perturbed state observations, which aligns with the setting of many adversarial attacks to deep reinforcement learning (DRL) and is also important for rolling out real-world RL agent under unpredictable sensing noise. With a fixed agent policy, we demonstrate that an optimal adversary to perturb state observations can be found, which is guaranteed to obtain the worst case agent reward. For DRL settings, this leads to a novel empirical adversarial attack to RL agents via a learned adversary that is much stronger than previous ones. To enhance the robustness of an agent, we propose a framework of alternating training with learned adversaries (ATLA), which trains an adversary online together with the agent using policy gradient following the optimal adversarial attack framework. Additionally, inspired by the analysis of state-adversarial Markov decision process (SA-MDP), we show that past states and actions (history) can be useful for learning a robust agent, and we empirically find a LSTM based policy can be more robust under adversaries. Empirical evaluations on a few continuous control environments show that ATLA achieves state-of-the-art performance under strong adversaries. Our code is available at https://github.com/huanzhang12/ATLA_robust_RL.