Many hierarchical reinforcement learning (RL) applications have empirically verified that incorporating prior knowledge in reward design improves convergence speed and practical performance. We attempt to quantify the computational benefits of hierarchical RL from a planning perspective under assumptions about the intermediate state and intermediate rewards frequently (but often implicitly) adopted in practice. Our approach reveals a trade-off between computational complexity and the pursuit of the shortest path in hierarchical planning: using intermediate rewards significantly reduces the computational complexity in finding a successful policy but does not guarantee to find the shortest path, whereas using sparse terminal rewards finds the shortest path at a significantly higher computational cost. We also corroborate our theoretical results with extensive experiments on the MiniGrid environments using Q-learning and other popular deep RL algorithms.
One of the most widely used methods for solving large-scale stochastic optimization problems is distributed asynchronous stochastic gradient descent (DASGD), a family of algorithms that result from parallelizing stochastic gradient descent on distributed computing architectures (possibly) asychronously. However, a key obstacle in the efficient implementation of DASGD is the issue of delays: when a computing node contributes a gradient update, the global model parameter may have already been updated by other nodes several times over, thereby rendering this gradient information stale. These delays can quickly add up if the computational throughput of a node is saturated, so the convergence of DASGD may be compromised in the presence of large delays. Our first contribution is that, by carefully tuning the algorithm's step-size, convergence to the critical set is still achieved in mean square, even if the delays grow unbounded at a polynomial rate. We also establish finer results in a broad class of structured optimization problems (called variationally coherent), where we show that DASGD converges to a global optimum with probability $1$ under the same delay assumptions. Together, these results contribute to the broad landscape of large-scale non-convex stochastic optimization by offering state-of-the-art theoretical guarantees and providing insights for algorithm design.
Learning optimal policies from historical data enables the gains from personalization to be realized in a wide variety of applications. The growing policy learning literature focuses on a setting where the treatment assignment policy does not adapt to the data. However, adaptive data collection is becoming more common in practice, from two primary sources: 1) data collected from adaptive experiments that are designed to improve inferential efficiency; 2) data collected from production systems that are adaptively evolving an operational policy to improve performance over time (e.g. contextual bandits). In this paper, we aim to address the challenge of learning the optimal policy with adaptively collected data and provide one of the first theoretical inquiries into this problem. We propose an algorithm based on generalized augmented inverse propensity weighted estimators and establish its finite-sample regret bound. We complement this regret upper bound with a lower bound that characterizes the fundamental difficulty of policy learning with adaptive data. Finally, we demonstrate our algorithm's effectiveness using both synthetic data and public benchmark datasets.
We design a simple reinforcement learning agent that, with a specification only of agent state dynamics and a reward function, can operate with some degree of competence in any environment. The agent maintains only visitation counts and value estimates for each agent-state-action pair. The value function is updated incrementally in response to temporal differences and optimistic boosts that encourage exploration. The agent executes actions that are greedy with respect to this value function. We establish a regret bound demonstrating convergence to near-optimal per-period performance, where the time taken to achieve near-optimality is polynomial in the number of agent states and actions, as well as the reward mixing time of the best policy within the reference policy class, which is comprised of those that depend on history only through agent state. Notably, there is no further dependence on the number of environment states or mixing times associated with other policies or statistics of history. Our result sheds light on the potential benefits of (deep) representation learning, which has demonstrated the capability to extract compact and relevant features from high-dimensional interaction histories.
Consider a player that in each round $t$ out of $T$ rounds chooses an action and observes the incurred cost after a delay of $d_{t}$ rounds. The cost functions and the delay sequence are chosen by an adversary. We show that even if the players' algorithms lose their "no regret" property due to too large delays, the expected discounted ergodic distribution of play converges to the set of coarse correlated equilibrium (CCE) if the algorithms have "no discounted-regret". For a zero-sum game, we show that no discounted-regret is sufficient for the discounted ergodic average of play to converge to the set of Nash equilibria. We prove that the FKM algorithm with $n$ dimensions achieves a regret of $O\left(nT^{\frac{3}{4}}+\sqrt{n}T^{\frac{1}{3}}D^{\frac{1}{3}}\right)$ and the EXP3 algorithm with $K$ arms achieves a regret of $O\left(\sqrt{\ln K\left(KT+D\right)}\right)$ even when $D=\sum_{t=1}^{T}d_{t}$ and $T$ are unknown. These bounds use a novel doubling trick that provably retains the regret bound for when $D$ and $T$ are known. Using these bounds, we show that EXP3 and FKM have no discounted-regret even for $d_{t}=O\left(t\log t\right)$. Therefore, the CCE of a finite or convex unknown game can be approximated even when only delayed bandit feedback is available via simulation.
To balance exploration and exploitation, multi-armed bandit algorithms need to conduct inference on the true mean reward of each arm in every time step using the data collected so far. However, the history of arms and rewards observed up to that time step is adaptively collected and there are known challenges in conducting inference with non-iid data. In particular, sample averages, which play a prominent role in traditional upper confidence bound algorithms and traditional Thompson sampling algorithms, are neither unbiased nor asymptotically normal. We propose a variant of a Thompson sampling based algorithm that leverages recent advances in the causal inference literature and adaptively re-weighs the terms of a doubly robust estimator on the true mean reward of each arm -- hence its name doubly-adaptive Thompson sampling. The regret of the proposed algorithm matches the optimal (minimax) regret rate and its empirical evaluation in a semi-synthetic experiment based on data from a randomized control trial of a web service is performed: we see that the proposed doubly-adaptive Thompson sampling has superior empirical performance to existing baselines in terms of cumulative regret and statistical power in identifying the best arm. Further, we extend this approach to contextual bandits, where there are more sources of bias present apart from the adaptive data collection -- such as the mismatch between the true data generating process and the reward model assumptions or the unequal representations of certain regions of the context space in initial stages of learning -- and propose the linear contextual doubly-adaptive Thompson sampling and the non-parametric contextual doubly-adaptive Thompson sampling extensions of our approach.
In many multi-agent reinforcement learning applications such as flocking, multi-robot applications and smart manufacturing, distinct agents share similar dynamics but face different objectives. In these applications, an important question is how the similarities amongst the agents can accelerate learning in spite of the agents' differing goals. We study a distributed LQR (Linear Quadratic Regulator) tracking problem which models this setting, where the agents, acting independently, share identical (unknown) dynamics and cost structure but need to track different targets. In this paper, we propose a communication-efficient, federated model-free zeroth-order algorithm that provably achieves a convergence speedup linear in the number of agents compared with the communication-free setup where each agent's problem is treated independently. We support our arguments with numerical simulations of both linear and nonlinear systems.
We study the problem of dynamic batch learning in high-dimensional sparse linear contextual bandits, where a decision maker, under a given maximum-number-of-batch constraint and only able to observe rewards at the end of each batch, can dynamically decide how many individuals to include in the next batch (at the end of the current batch) and what personalized action-selection scheme to adopt within each batch. Such batch constraints are ubiquitous in a variety of practical contexts, including personalized product offerings in marketing and medical treatment selection in clinical trials. We characterize the fundamental learning limit in this problem via a regret lower bound and provide a matching upper bound (up to log factors), thus prescribing an optimal scheme for this problem. To the best of our knowledge, our work provides the first inroad into a theoretical understanding of dynamic batch learning in high-dimensional sparse linear contextual bandits. Notably, even a special case of our result (when no batch constraint is present) yields the first minimax optimal $\tilde{O}(\sqrt{s_0T})$ regret bound for standard online learning in high-dimensional linear contextual bandits (for the no-margin case), where $s_0$ is the sparsity parameter (or an upper bound thereof) and $T$ is the learning horizon. This result (both that $\tilde{O}(\sqrt{s_0 T})$ is achievable and that $\Omega(\sqrt{s_0 T})$ is a lower bound) appears to be unknown in the emerging literature of high-dimensional contextual bandits.