In a single-agent setting, reinforcement learning (RL) tasks can be cast into an inference problem by introducing a binary random variable o, which stands for the "optimality". In this paper, we redefine the binary random variable o in multi-agent setting and formalize multi-agent reinforcement learning (MARL) as probabilistic inference. We derive a variational lower bound of the likelihood of achieving the optimality and name it as Regularized Opponent Model with Maximum Entropy Objective (ROMMEO). From ROMMEO, we present a novel perspective on opponent modeling and show how it can improve the performance of training agents theoretically and empirically in cooperative games. To optimize ROMMEO, we first introduce a tabular Q-iteration method ROMMEO-Q with proof of convergence. We extend the exact algorithm to complex environments by proposing an approximate version, ROMMEO-AC. We evaluate these two algorithms on the challenging iterated matrix game and differential game respectively and show that they can outperform strong MARL baselines.

We describe a new approach for mitigating risk in the Reinforcement Learning paradigm. Instead of reasoning about expected utility, we use second-order stochastic dominance (SSD) to directly compare the inherent risk of random returns induced by different actions. We frame the RL optimization within the space of probability measures to accommodate the SSD relation, treating Bellman's equation as a potential energy functional. This brings us to Wasserstein gradient flows, for which the optimality and convergence are well understood. We propose a discrete-measure approximation algorithm called the Dominant Particle Agent (DPA), and we demonstrate how safety and performance are better balanced with DPA than with existing baselines.

Shaping in humans and animals has been shown to be a powerful tool for learning complex tasks as compared to learning in a randomized fashion. This makes the problem less complex and enables one to solve the easier sub task at hand first. Generating a curriculum for such guided learning involves subjecting the agent to easier goals first, and then gradually increasing their difficulty. This paper takes a similar direction and proposes a dual curriculum scheme for solving robotic manipulation tasks with sparse rewards, called MaMiC. It includes a macro curriculum scheme which divides the task into multiple sub-tasks followed by a micro curriculum scheme which enables the agent to learn between such discovered sub-tasks. We show how combining macro and micro curriculum strategies help in overcoming major exploratory constraints considered in robot manipulation tasks without having to engineer any complex rewards. We also illustrate the meaning of the individual curricula and how they can be used independently based on the task. The performance of such a dual curriculum scheme is analyzed on the Fetch environments.

As global greenhouse gas emissions continue to rise, the use of stratospheric aerosol injection (SAI), a form of solar geoengineering, is increasingly considered in order to artificially mitigate climate change effects. However, initial research in simulation suggests that naive SAI can have catastrophic regional consequences, which may induce serious geostrategic conflicts. Current geo-engineering research treats SAI control in low-dimensional approximation only. We suggest treating SAI as a high-dimensional control problem, with policies trained according to a context-sensitive reward function within the Deep Reinforcement Learning (DRL) paradigm. In order to facilitate training in simulation, we suggest to emulate HadCM3, a widely used General Circulation Model, using deep learning techniques. We believe this is the first application of DRL to the climate sciences.

We assume that we are given a time series of data from a dynamical system and our task is to learn the flow map of the dynamical system. We present a collection of results on how to enforce constraints coming from the dynamical system in order to accelerate the training of deep neural networks to represent the flow map of the system as well as increase their predictive ability. In particular, we provide ways to enforce constraints during training for all three major modes of learning, namely supervised, unsupervised and reinforcement learning. In general, the dynamic constraints need to include terms which are analogous to memory terms in model reduction formalisms. Such memory terms act as a restoring force which corrects the errors committed by the learned flow map during prediction. For supervised learning, the constraints are added to the objective function. For the case of unsupervised learning, in particular generative adversarial networks, the constraints are introduced by augmenting the input of the discriminator. Finally, for the case of reinforcement learning and in particular actor-critic methods, the constraints are added to the reward function. In addition, for the reinforcement learning case, we present a novel approach based on homotopy of the action-value function in order to stabilize and accelerate training. We use numerical results for the Lorenz system to illustrate the various constructions.

Off-policy reinforcement learning with eligibility traces is challenging because of the discrepancy between target policy and behavior policy. One common approach is to measure the difference between two policies in a probabilistic way, such as importance sampling and tree-backup. However, existing off-policy learning methods based on probabilistic policy measurement are inefficient when utilizing traces under a greedy target policy, which is ineffective for control problems. The traces are cut immediately when a non-greedy action is taken, which may lose the advantage of eligibility traces and slow down the learning process. Alternatively, some non-probabilistic measurement methods such as General Q($\lambda$) and Naive Q($\lambda$) never cut traces, but face convergence problems in practice. To address the above issues, this paper introduces a new method named TBQ($\sigma$), which effectively unifies the tree-backup algorithm and Naive Q($\lambda$). By introducing a new parameter $\sigma$ to illustrate the \emph{degree} of utilizing traces, TBQ($\sigma$) creates an effective integration of TB($\lambda$) and Naive Q($\lambda$) and continuous role shift between them. The contraction property of TB($\sigma$) is theoretically analyzed for both policy evaluation and control settings. We also derive the online version of TBQ($\sigma$) and give the convergence proof. We empirically show that, for $\epsilon\in(0,1]$ in $\epsilon$-greedy policies, there exists some degree of utilizing traces for $\lambda\in[0,1]$, which can improve the efficiency in trace utilization for off-policy reinforcement learning, to both accelerate the learning process and improve the performance.

Recent work in reinforcement learning demonstrated that learning solely through self-play is not only possible, but could also result in novel strategies that humans never would have thought of. However, optimization methods cast as a game between two players require careful tuning to prevent suboptimal results. Hence, we look at random play as an alternative method. In this paper, we train a DQN agent to play Sungka, a two-player turn-based board game wherein the players compete to obtain more stones than the other. We show that even with purely random play, our training algorithm converges very fast and is stable. Moreover, we test our trained agent against several baselines and show its ability to consistently win against these.

Deep reinforcement learning (DRL) algorithms have achieved great success on sequential decision-making problems, yet is criticized for the lack of data-efficiency and explainability. Especially, explainability of subtasks is critical in hierarchical decision-making since it enhances the transparency of black-box-style DRL methods and helps the RL practitioners to understand the high-level behavior of the system better. To improve the data-efficiency and explainability of DRL, declarative knowledge is introduced in this work and novel algorithm is proposed by integrating DRL with symbolic planning. An experimental analysis on publicly available benchmarks validates the explainability of the subtasks and shows that our method can outperform the state-of-the-art approach in terms of data-efficiency.

Exploration is a crucial component for discovering approximately optimal policies in most high-dimensional reinforcement learning (RL) settings with sparse rewards. Approaches such as neural density models and continuous exploration (e.g., Go-Explore) have been instrumental in recent advances. Soft actor-critic (SAC) is a method for improving exploration that aims to combine off-policy updates while maximizing the policy entropy. We extend SAC to a richer class of probability distributions through normalizing flows, which we show improves performance in exploration, sample complexity, and convergence. Finally, we show that not only the normalizing flow policy outperforms SAC on MuJoCo domains, it is also significantly lighter, using as low as 5.6% of the original network's parameters for similar performance.

Penetration testing (pentesting) involves performing a controlled attack on a computer system in order to assess it's security. Although an effective method for testing security, pentesting requires highly skilled practitioners and currently there is a growing shortage of skilled cyber security professionals. One avenue for alleviating this problem is automate the pentesting process using artificial intelligence techniques. Current approaches to automated pentesting have relied on model-based planning, however the cyber security landscape is rapidly changing making maintaining up-to-date models of exploits a challenge. This project investigated the application of model-free Reinforcement Learning (RL) to automated pentesting. Model-free RL has the key advantage over model-based planning of not requiring a model of the environment, instead learning the best policy through interaction with the environment. We first designed and built a fast, low compute simulator for training and testing autonomous pentesting agents. We did this by framing pentesting as a Markov Decision Process with the known configuration of the network as states, the available scans and exploits as actions, the reward determined by the value of machines on the network. We then used this simulator to investigate the application of model-free RL to pentesting. We tested the standard Q-learning algorithm using both tabular and neural network based implementations. We found that within the simulated environment both tabular and neural network implementations were able to find optimal attack paths for a range of different network topologies and sizes without having a model of action behaviour. However, the implemented algorithms were only practical for smaller networks and numbers of actions. Further work is needed in developing scalable RL algorithms and testing these algorithms in larger and higher fidelity environments.