Posted by Pablo Samuel Castro, Staff Software Engineer, Google Research It is widely accepted that the enormous growth of deep reinforcem...

A dynamic treatment regimen (DTR) is a set of decision rules to personalize treatments for an individual using their medical history. The Q-learning based Q-shared algorithm has been used to develop DTRs that involve decision rules shared across multiple stages of intervention. We show that the existing Q-shared algorithm can suffer from non-convergence due to the use of linear models in the Q-learning setup, and identify the condition in which Q-shared fails. Leveraging properties from expansion-constrained ordinary least-squares, we give a penalized Q-shared algorithm that not only converges in settings that violate the condition, but can outperform the original Q-shared algorithm even when the condition is satisfied. We give evidence for the proposed method in a real-world application and several synthetic simulations.

Reinforcement learning (RL) in partially observable, fully cooperative multi-agent settings (Dec-POMDPs) can in principle be used to address many real-world challenges such as controlling a swarm of rescue robots or a synchronous team of quadcopters. However, Dec-POMDPs are significantly harder to solve than single-agent problems, with the former being NEXP-complete and the latter, MDPs, being just P-complete. Hence, current RL algorithms for Dec-POMDPs suffer from poor sample complexity, thereby reducing their applicability to practical problems where environment interaction is costly. Our key insight is that using just a polynomial number of samples, one can learn a centralized model that generalizes across different policies. We can then optimize the policy within the learned model instead of the true system, reducing the number of environment interactions. We also learn a centralized exploration policy within our model that learns to collect additional data in state-action regions with high model uncertainty. Finally, we empirically evaluate the proposed model-based algorithm, MARCO, in three cooperative communication tasks, where it improves sample efficiency by up to 20x.

We propose the k-Shortest-Path (k-SP) constraint: a novel constraint on the agent's trajectory that improves the sample efficiency in sparse-reward MDPs. We show that any optimal policy necessarily satisfies the k-SP constraint. Notably, the k-SP constraint prevents the policy from exploring state-action pairs along the non-k-SP trajectories (e.g., going back and forth). However, in practice, excluding state-action pairs may hinder the convergence of RL algorithms. To overcome this, we propose a novel cost function that penalizes the policy violating SP constraint, instead of completely excluding it. Our numerical experiment in a tabular RL setting demonstrates that the SP constraint can significantly reduce the trajectory space of policy. As a result, our constraint enables more sample efficient learning by suppressing redundant exploration and exploitation. Our experiments on MiniGrid, DeepMind Lab, Atari, and Fetch show that the proposed method significantly improves proximal policy optimization (PPO) and outperforms existing novelty-seeking exploration methods including count-based exploration even in continuous control tasks, indicating that it improves the sample efficiency by preventing the agent from taking redundant actions.

We study model-based offline Reinforcement Learning with general function approximation. We present an algorithm named Constrained Pessimistic Policy Optimization (CPPO) which leverages a general function class and uses a constraint to encode pessimism. Under the assumption that the ground truth model belongs to our function class, CPPO can learn with the offline data only providing partial coverage, i.e., it can learn a policy that completes against any policy that is covered by the offline data, in polynomial sample complexity with respect to the statistical complexity of the function class. We then demonstrate that this algorithmic framework can be applied to many specialized Markov Decision Processes where the additional structural assumptions can further refine the concept of partial coverage. One notable example is low-rank MDP with representation learning where the partial coverage is defined using the concept of relative condition number measured by the underlying unknown ground truth feature representation. Finally, we introduce and study the Bayesian setting in offline RL. The key benefit of Bayesian offline RL is that algorithmically, we do not need to explicitly construct pessimism or reward penalty which could be hard beyond models with linear structures. We present a posterior sampling-based incremental policy optimization algorithm (PS-PO) which proceeds by iteratively sampling a model from the posterior distribution and performing one-step incremental policy optimization inside the sampled model. Theoretically, in expectation with respect to the prior distribution, PS-PO can learn a near optimal policy under partial coverage with polynomial sample complexity.

Inefficient traffic signal control methods may cause numerous problems, such as traffic congestion and waste of energy. Reinforcement learning (RL) is a trending data-driven approach for adaptive traffic signal control in complex urban traffic networks. Although the development of deep neural networks (DNN) further enhances its learning capability, there are still some challenges in applying deep RLs to transportation networks with multiple signalized intersections, including non-stationarity environment, exploration-exploitation dilemma, multi-agent training schemes, continuous action spaces, etc. In order to address these issues, this paper first proposes a multi-agent deep deterministic policy gradient (MADDPG) method by extending the actor-critic policy gradient algorithms. MADDPG has a centralized learning and decentralized execution paradigm in which critics use additional information to streamline the training process, while actors act on their own local observations. The model is evaluated via simulation on the Simulation of Urban MObility (SUMO) platform. Model comparison results show the efficiency of the proposed algorithm in controlling traffic lights.

As the number of Human-Centered Internet of Things (HCIoT) applications increases, the self-adaptation of its services and devices is becoming a fundamental requirement for addressing the uncertainties of the environment in decision-making processes. Self-adaptation of HCIoT aims to manage run-time changes in a dynamic environment and to adjust the functionality of IoT objects in order to achieve desired goals during execution. SMASH is a semantic-enabled multi-agent system for self-adaptation of HCIoT that autonomously adapts IoT objects to uncertainties of their environment. SMASH addresses the self-adaptation of IoT applications only according to the human values of users, while the behavior of users is not addressed. This article presents Q-SMASH: a multi-agent reinforcement learning-based approach for self-adaptation of IoT objects in human-centered environments. Q-SMASH aims to learn the behaviors of users along with respecting human values. The learning ability of Q-SMASH allows it to adapt itself to the behavioral change of users and make more accurate decisions in different states and situations.

What's the best way to arrange wells in an oil or gas field? It's a simple enough question, but the answer can be very complex. Now a Cal Tech/JPL spinoff is developing a new approach that blends traditional HPC simulation with deep reinforcement learning running on GPUs to optimize energy extraction. The well placement game is a familiar one to oil and gas companies. For years, they have been using simulators running atop HPC systems to model underground reservoirs.

Reinforcement learning (RL) provides a framework for learning goal-directed policies given user-specified rewards. However, since designing rewards often requires substantial engineering effort, we are interested in the problem of learning without rewards, where agents must discover useful behaviors in the absence of task-specific incentives. Intrinsic motivation is a family of unsupervised RL techniques which develop general objectives for an RL agent to optimize that lead to better exploration or the discovery of skills. In this paper, we propose a new unsupervised RL technique based on an adversarial game which pits two policies against each other to compete over the amount of surprise an RL agent experiences. The policies each take turns controlling the agent. The Explore policy maximizes entropy, putting the agent into surprising or unfamiliar situations. Then, the Control policy takes over and seeks to recover from those situations by minimizing entropy. The game harnesses the power of multi-agent competition to drive the agent to seek out increasingly surprising parts of the environment while learning to gain mastery over them. We show empirically that our method leads to the emergence of complex skills by exhibiting clear phase transitions. Furthermore, we show both theoretically (via a latent state space coverage argument) and empirically that our method has the potential to be applied to the exploration of stochastic, partially-observed environments. We show that Adversarial Surprise learns more complex behaviors, and explores more effectively than competitive baselines, outperforming intrinsic motivation methods based on active inference, novelty-seeking (Random Network Distillation (RND)), and multi-agent unsupervised RL (Asymmetric Self-Play (ASP)) in MiniGrid, Atari and VizDoom environments.

Learning data representations that are useful for various downstream tasks is a cornerstone of artificial intelligence. While existing methods are typically evaluated on downstream tasks such as classification or generative image quality, we propose to assess representations through their usefulness in downstream control tasks, such as reaching or pushing objects. By training over 10,000 reinforcement learning policies, we extensively evaluate to what extent different representation properties affect out-of-distribution (OOD) generalization. Finally, we demonstrate zero-shot transfer of these policies from simulation to the real world, without any domain randomization or fine-tuning. This paper aims to establish the first systematic characterization of the usefulness of learned representations for real-world OOD downstream tasks.