We study optimality for the safety-constrained Markov decision process which is the underlying framework for safe reinforcement learning. Specifically, we consider a constrained Markov decision process (with finite states and finite actions) where the goal of the decision maker is to reach a target set while avoiding an unsafe set(s) with certain probabilistic guarantees. Therefore the underlying Markov chain for any control policy will be multichain since by definition there exists a target set and an unsafe set. The decision maker also has to be optimal (with respect to a cost function) while navigating to the target set. This gives rise to a multi-objective optimization problem. We highlight the fact that Bellman's principle of optimality may not hold for constrained Markov decision problems with an underlying multichain structure (as shown by the counterexample due to Haviv. We resolve the counterexample by formulating the aforementioned multi-objective optimization problem as a zero-sum game and thereafter construct an asynchronous value iteration scheme for the Lagrangian (similar to Shapley's algorithm). Finally, we consider the reinforcement learning problem for the same and construct a modified $Q$-learning algorithm for learning the Lagrangian from data. We also provide a lower bound on the number of iterations required for learning the Lagrangian and corresponding error bounds.

Temporal difference (TD) learning is a policy evaluation in reinforcement learning whose performance can be enhanced by variance reduction techniques. Recently, multiple works have sought to fuse TD learning with SVRG to obtain a policy evaluation method with a geometric rate of convergence. However, the resulting convergence rate is significantly weaker than what is achieved by SVRG in the setting of convex optimization. In this work we utilize a recent interpretation of TD-learning as the splitting of the gradient of an appropriately chosen function, thus simplifying the algorithm and fusing TD with SVRG. Our main result is a geometric convergence bound with predetermined learning rate of $1/8$, which is identical to the convergence bound available for SVRG in the convex setting. Our theoretical findings are supported by a set of experiments.

A dynamic mean field theory is developed for finite state and action Bayesian reinforcement learning in the large state space limit. In an analogy with statistical physics, the Bellman equation is studied as a disordered dynamical system; the Markov decision process transition probabilities are interpreted as couplings and the value functions as deterministic spins that evolve dynamically. Thus, the mean-rewards and transition probabilities are considered to be quenched random variables. The theory reveals that, under certain assumptions, the state-action values are statistically independent across state-action pairs in the asymptotic state space limit, and provides the form of the distribution exactly. The results hold in the finite and discounted infinite horizon settings, for both value iteration and policy evaluation. The state-action value statistics can be computed from a set of mean field equations, which we call dynamic mean field programming (DMFP). For policy evaluation the equations are exact. For value iteration, approximate equations are obtained by appealing to extreme value theory or bounds. The result provides analytic insight into the statistical structure of tabular reinforcement learning, for example revealing the conditions under which reinforcement learning is equivalent to a set of independent multi-armed bandit problems.

In this thesis, we explore the use of complex systems to study learning and adaptation in natural and artificial systems. The goal is to develop autonomous systems that can learn without supervision, develop on their own, and become increasingly complex over time. Complex systems are identified as a suitable framework for understanding these phenomena due to their ability to exhibit growth of complexity. Being able to build learning algorithms that require limited to no supervision would enable greater flexibility and adaptability in various applications. By understanding the fundamental principles of learning in complex systems, we hope to advance our ability to design and implement practical learning algorithms in the future. This thesis makes the following key contributions: the development of a general complexity metric that we apply to search for complex systems that exhibit growth of complexity, the introduction of a coarse-graining method to study computations in large-scale complex systems, and the development of a metric for learning efficiency as well as a benchmark dataset for evaluating the speed of learning algorithms. Our findings add substantially to our understanding of learning and adaptation in natural and artificial systems. Moreover, our approach contributes to a promising new direction for research in this area. We hope these findings will inspire the development of more effective and efficient learning algorithms in the future.

Recently, DARPA launched the ShELL program, which aims to explore how experience sharing can benefit distributed lifelong learning agents in adapting to new challenges. In this paper, we address this issue by conducting both theoretical and empirical research on distributed multi-task reinforcement learning (RL), where a group of $N$ agents collaboratively solves $M$ tasks without prior knowledge of their identities. We approach the problem by formulating it as linearly parameterized contextual Markov decision processes (MDPs), where each task is represented by a context that specifies the transition dynamics and rewards. To tackle this problem, we propose an algorithm called DistMT-LSVI. First, the agents identify the tasks, and then they exchange information through a central server to derive $\epsilon$-optimal policies for the tasks. Our research demonstrates that to achieve $\epsilon$-optimal policies for all $M$ tasks, a single agent using DistMT-LSVI needs to run a total number of episodes that is at most $\tilde{\mathcal{O}}({d^3H^6(\epsilon^{-2}+c_{\rm sep}^{-2})}\cdot M/N)$, where $c_{\rm sep}>0$ is a constant representing task separability, $H$ is the horizon of each episode, and $d$ is the feature dimension of the dynamics and rewards. Notably, DistMT-LSVI improves the sample complexity of non-distributed settings by a factor of $1/N$, as each agent independently learns $\epsilon$-optimal policies for all $M$ tasks using $\tilde{\mathcal{O}}(d^3H^6M\epsilon^{-2})$ episodes. Additionally, we provide numerical experiments conducted on OpenAI Gym Atari environments that validate our theoretical findings.

Due to technological advances in the field of radio technology and its availability, the number of interference signals in the radio spectrum is continuously increasing. Interference signals must be detected in a timely fashion, in order to maintain standards and keep emergency frequencies open. To this end, specialized (multi-channel) receivers are used for spectrum monitoring. In this paper, the performances of two different approaches for controlling the available receiver resources are compared. The methods used for resource management (ReMa) are linear frequency tuning as a heuristic approach and a Q-learning algorithm from the field of reinforcement learning. To test the methods to be investigated, a simplified scenario was designed with two receiver channels monitoring ten non-overlapping frequency bands with non-uniform signal activity. For this setting, it is shown that the Q-learning algorithm used has a significantly higher detection rate than the heuristic approach at the expense of a smaller exploration rate. In particular, the Q-learning approach can be parameterized to allow for a suitable trade-off between detection and exploration rate.

Due to shortage of water resources and increasing water demands, the joint operation of multireservoir systems for balancing power generation, ecological protection, and the residential water supply has become a critical issue in hydropower management. However, the numerous constraints and nonlinearity of multiple reservoirs make solving this problem time-consuming. To address this challenge, a deep reinforcement learning approach that incorporates a transformer framework is proposed. The multihead attention mechanism of the encoder effectively extracts information from reservoirs and residential areas, and the multireservoir attention network of the decoder generates suitable operational decisions. The proposed method is applied to Lake Mead and Lake Powell in the Colorado River Basin. The experimental results demonstrate that the transformer-based deep reinforcement learning approach can produce appropriate operational outcomes. Compared to a state-of-the-art method, the operation strategies produced by the proposed approach generate 10.11% more electricity, reduce the amended annual proportional flow deviation by 39.69%, and increase water supply revenue by 4.10%. Consequently, the proposed approach offers an effective method for the multiobjective operation of multihydropower reservoir systems.

Despite their unprecedented success, even the largest language models make mistakes. Similar to how humans learn and improve using feedback, previous work proposed providing language models with natural language feedback to guide them in repairing their outputs. Because human-generated critiques are expensive to obtain, researchers have devised learned critique generators in lieu of human critics while assuming one can train downstream models to utilize generated feedback. However, this approach does not apply to black-box or limited access models such as ChatGPT, as they cannot be fine-tuned. Moreover, in the era of large general-purpose language agents, fine-tuning is neither computationally nor spatially efficient as it results in multiple copies of the network. In this work, we introduce RL4F (Reinforcement Learning for Feedback), a multi-agent collaborative framework where the critique generator is trained to maximize end-task performance of GPT-3, a fixed model more than 200 times its size. RL4F produces critiques that help GPT-3 revise its outputs. We study three datasets for action planning, summarization and alphabetization and show relative improvements up to 10% in multiple text similarity metrics over other learned, retrieval-augmented or prompting-based critique generators.

Adversarial attacks in reinforcement learning (RL) often assume highly-privileged access to the victim's parameters, environment, or data. Instead, this paper proposes a novel adversarial setting called a Cheap Talk MDP in which an Adversary can merely append deterministic messages to the Victim's observation, resulting in a minimal range of influence. The Adversary cannot occlude ground truth, influence underlying environment dynamics or reward signals, introduce non-stationarity, add stochasticity, see the Victim's actions, or access their parameters. Additionally, we present a simple meta-learning algorithm called Adversarial Cheap Talk (ACT) to train Adversaries in this setting. We demonstrate that an Adversary trained with ACT still significantly influences the Victim's training and testing performance, despite the highly constrained setting. Affecting train-time performance reveals a new attack vector and provides insight into the success and failure modes of existing RL algorithms. More specifically, we show that an ACT Adversary is capable of harming performance by interfering with the learner's function approximation, or instead helping the Victim's performance by outputting useful features. Finally, we show that an ACT Adversary can manipulate messages during train-time to directly and arbitrarily control the Victim at test-time. Project video and code are available at https://sites.google.com/view/adversarial-cheap-talk

Model-Based Reinforcement Learning (RL) is widely believed to have the potential to improve sample efficiency by allowing an agent to synthesize large amounts of imagined experience. Experience Replay (ER) can be considered a simple kind of model, which has proved effective at improving the stability and efficiency of deep RL. In principle, a learned parametric model could improve on ER by generalizing from real experience to augment the dataset with additional plausible experience. However, given that learned value functions can also generalize, it is not immediately obvious why model generalization should be better. Here, we provide theoretical and empirical insight into when, and how, we can expect data generated by a learned model to be useful. First, we provide a simple theorem motivating how learning a model as an intermediate step can narrow down the set of possible value functions more than learning a value function directly from data using the Bellman equation. Second, we provide an illustrative example showing empirically how a similar effect occurs in a more concrete setting with neural network function approximation. Finally, we provide extensive experiments showing the benefit of model-based learning for online RL in environments with combinatorial complexity, but factored structure that allows a learned model to generalize. In these experiments, we take care to control for other factors in order to isolate, insofar as possible, the benefit of using experience generated by a learned model relative to ER alone.