Microfluidic devices are utilized to control and direct flow behavior in a wide variety of applications, particularly in medical diagnostics. A particularly popular form of microfluidics -- called inertial microfluidic flow sculpting -- involves placing a sequence of pillars to controllably deform an initial flow field into a desired one. Inertial flow sculpting can be formally defined as an inverse problem, where one identifies a sequence of pillars (chosen, with replacement, from a finite set of pillars, each of which produce a specific transformation) whose composite transformation results in a user-defined desired transformation. Endemic to most such problems in engineering, inverse problems are usually quite computationally intractable, with most traditional approaches based on search and optimization strategies. In this paper, we pose this inverse problem as a Reinforcement Learning (RL) problem. We train a DoubleDQN agent to learn from this environment. The results suggest that learning is possible using a DoubleDQN model with the success frequency reaching 90% in 200,000 episodes and the rewards converging. While most of the results are obtained by fixing a particular target flow shape to simplify the learning problem, we later demonstrate how to transfer the learning of an agent based on one target shape to another, i.e. from one design to another and thus be useful for a generic design of a flow shape.

In this empirical paper, we investigate how learning agents can be arranged in more efficient communication topologies for improved learning. This is an important problem because a common technique to improve speed and robustness of learning in deep reinforcement learning and many other machine learning algorithms is to run multiple learning agents in parallel. The standard communication architecture typically involves all agents intermittently communicating with each other (fully connected topology) or with a centralized server (star topology). Unfortunately, optimizing the topology of communication over the space of all possible graphs is a hard problem, so we borrow results from the networked optimization and collective intelligence literatures which suggest that certain families of network topologies can lead to strong improvements over fully-connected networks. We start by introducing alternative network topologies to DRL benchmark tasks under the Evolution Strategies paradigm which we call Network Evolution Strategies. We explore the relative performance of the four main graph families and observe that one such family (Erdos-Renyi random graphs) empirically outperforms all other families, including the de facto fully-connected communication topologies. Additionally, the use of alternative network topologies has a multiplicative performance effect: we observe that when 1000 learning agents are arranged in a carefully designed communication topology, they can compete with 3000 agents arranged in the de facto fully-connected topology. Overall, our work suggests that distributed machine learning algorithms would learn more efficiently if the communication topology between learning agents was optimized.

In reinforcement learning, the reward function on current state and action is widely used. When the objective is about the expectation of the (discounted) total reward only, it works perfectly. However, if the objective involves the total reward distribution, the result will be wrong. This paper studies Value-at-Risk (VaR) problems in short- and long-horizon Markov decision processes (MDPs) with two reward functions, which share the same expectations. Firstly we show that with VaR objective, when the real reward function is transition-based (with respect to action and both current and next states), the simplified (state-based, with respect to action and current state only) reward function will change the VaR. Secondly, for long-horizon MDPs, we estimate the VaR function with the aid of spectral theory and the central limit theorem. Thirdly, since the estimation method is for a Markov reward process with the reward function on current state only, we present a transformation algorithm for the Markov reward process with the reward function on current and next states, in order to estimate the VaR function with an intact total reward distribution.

Two years ago, Alphabet researchers made computing history when their artificial intelligence software AlphaGo defeated a world champion at the complex board game Go. Amazon now hopes to democratize the AI technique behind that milestone--with a pint-size self-driving car. The 1/18th-scale vehicle is called DeepRacer, and it can be preordered for $249; it will later cost $399. It's designed to make it easier for programmers to get started with reinforcement learning, the technique that powered AlphaGo's victory and is loosely inspired by how animals learn from feedback on their behavior. Although the approach has produced notable research stunts, such as bots that can play Go, chess, and complicated multiplayer electronic games, it isn't as widely used as the pattern-matching learning techniques used in speech recognition and image analysis.

To overcome the curse of dimensionality and curse of modeling in Dynamic Programming (DP) methods for solving classical Markov Decision Process (MDP) problems, Reinforcement Learning (RL) algorithms are popular. In this paper, we consider an infinite-horizon average reward MDP problem and prove the optimality of the threshold policy under certain conditions. Traditional RL techniques do not exploit the threshold nature of optimal policy while learning. In this paper, we propose a new RL algorithm which utilizes the known threshold structure of the optimal policy while learning by reducing the feasible policy space. We establish that the proposed algorithm converges to the optimal policy. It provides a significant improvement in convergence speed and computational and storage complexity over traditional RL algorithms. The proposed technique can be applied to a wide variety of optimization problems that include energy efficient data transmission and management of queues. We exhibit the improvement in convergence speed of the proposed algorithm over other RL algorithms through simulations.

Entropy regularization is commonly used to improve policy optimization in reinforcement learning. It is believed to help with exploration by encouraging the selection of more stochastic policies. In this work, we analyze this claim and, through new visualizations of the optimization landscape, we observe that incorporating entropy in policy optimization serves as a regularizer. We show that even with access to the exact gradient, policy optimization is difficult due to the geometry of the objective function. We qualitatively show that, in some environments, entropy regularization can make the optimization landscape smoother, thereby connecting local optima and enabling the use of larger learning rates. This manuscript presents new tools for understanding the underlying optimization landscape and highlights the challenge of designing general-purpose policy optimization algorithms in reinforcement learning.

Continual learning is the problem of learning new tasks or knowledge while protecting old knowledge and ideally generalizing from old experience to learn new tasks faster. Neural networks trained by stochastic gradient descent often degrade on old tasks when trained successively on new tasks with different data distributions. This phenomenon, referred to as catastrophic forgetting, is considered a major hurdle to learning with non-stationary data or sequences of new tasks, and prevents networks from continually accumulating knowledge and skills. We examine this issue in the context of reinforcement learning, in a setting where an agent is exposed to tasks in a sequence. Unlike most other work, we do not provide an explicit indication to the model of task boundaries, which is the most general circumstance for a learning agent exposed to continuous experience. While various methods to counteract catastrophic forgetting have recently been proposed, we explore a straightforward, general, and seemingly overlooked solution - that of using experience replay buffers for all past events - with a mixture of on- and off-policy learning, leveraging behavioral cloning. We show that this strategy can still learn new tasks quickly yet can substantially reduce catastrophic forgetting in both Atari and DMLab domains, even matching the performance of methods that require task identities. When buffer storage is constrained, we confirm that a simple mechanism for randomly discarding data allows a limited size buffer to perform almost as well as an unbounded one.

This paper investigates how to utilize different forms of human interaction to safely train autonomous systems in real-time by learning from both human demonstrations and interventions. We implement two components of the Cycle-of-Learning for Autonomous Systems, which is our framework for combining multiple modalities of human interaction. The current effort employs human demonstrations to teach a desired behavior via imitation learning, then leverages intervention data to correct for undesired behaviors produced by the imitation learner to teach novel tasks to an autonomous agent safely, after only minutes of training. We demonstrate this method in an autonomous perching task using a quadrotor with continuous roll, pitch, yaw, and throttle commands and imagery captured from a downward-facing camera in a high-fidelity simulated environment. Our method improves task completion performance for the same amount of human interaction when compared to learning from demonstrations alone, while also requiring on average 32% less data to achieve that performance. This provides evidence that combining multiple modes of human interaction can increase both the training speed and overall performance of policies for autonomous systems.

Being able to navigate to a target with minimal supervision and prior knowledge is critical to creating human-like assistive agents. Prior work on map-based and map-less approaches have limited generalizability. In this paper, we present a novel approach, Hybrid Asynchronous Universal Successor Representations (HAUSR), which overcomes the problem of generalizability to new goals by adapting recent work on Universal Successor Representations with Asynchronous Actor-Critic Agents. We show that the agent was able to successfully reach novel goals and we were able to quickly fine-tune the network for adapting to new scenes. This opens up novel application scenarios where intelligent agents could learn from and adapt to a wide range of environments with minimal human input.

An important challenge in drug discovery is to find molecules with desired chemical properties. While ultimate usefulness as a drug can only be determined in a laboratory or clinical context, that process is expensive, and it is thus advantageous to pre-select likely candidates in software. While deep learning has been extensively investigated for molecular graph encoding ([Duvenaud et al., 2015], [Kearnes et al., 2016], [Gilmer et al., 2017]), molecule generation is still subject of active research. The simplest natural approach to candidate molecule generation is to generate some sort of a linear representation,such as a string of characters in the SMILES format [Weininger, 1988], using an encoder-decoder network architecture similar to that used in machine translation, as done in [Gómez-Bombarelli et al., 2016]. This approach's performance was comparatively poor because a molecule's structure is not linear, but rather a graph which typically includes cycles, so it falls to the model to learn how to generate SMILES strings that correspond to chemically valid molecules - a nontrivial task that leaves the model with little spare capacity to additionally optimize a given chemical metric of the molecules produced. A way to partially remedy that involves generating not the actual SMILES strings, but a sequence ofproduction rules of a context-free grammar (CFG) for SMILES, as done by [Kusner et al., 2017]. That guarantees that the SMILES strings produced are grammatically valid, putting less burden on the model to ensure validity and thereby achieving better metrics. However, [Kusner et al., 2017] give two reasons why this is still not guaranteed to produce chemically valid molecules: firstly, a grammatically valid SMILES string is not guaranteed to be chemically possible (because of atom valences being wrong, for example), and secondly, because a