We consider the problem of offline reinforcement learning (RL) -- a well-motivated setting of RL that aims at policy optimization using only historical data. Despite its wide applicability, theoretical understandings of offline RL, such as its optimal sample complexity, remain largely open even in basic settings such as \emph{tabular} Markov Decision Processes (MDPs). In this paper, we propose Off-Policy Double Variance Reduction (OPDVR), a new variance reduction based algorithm for offline RL. Our main result shows that OPDVR provably identifies an ϵ-optimal policy with O˜(H2/dmϵ2) episodes of offline data in the finite-horizon stationary transition setting, where H is the horizon length and dm is the minimal marginal state-action distribution induced by the behavior policy. This improves over the best known upper bound by a factor of H. Moreover, we establish an information-theoretic lower bound of Ω(H2/dmϵ2) which certifies that OPDVR is optimal up to logarithmic factors. Lastly, we show that OPDVR also achieves rate-optimal sample complexity under alternative settings such as the finite-horizon MDPs with non-stationary transitions and the infinite horizon MDPs with discounted rewards.
more »
« less
Selfish Learning: Leveraging the Greed in Social Learning
We introduce a sequential Bayesian binary hypothesis testing problem under social learning, termed selfish learning, where agents work to maximize their individual rewards. In particular, each agent receives a private signal and is aware of decisions made by earlier-acting agents. Beside inferring the underlying hypothesis, agents also decide whether to stop and declare, or pass the inference to the next agent. The employer rewards only correct responses and the reward per worker decreases with the number of employees used for decision making. We characterize decision regions of agents in the infinite and finite horizon. In particular, we show that the decision boundaries in the infinite horizon are the solutions to a Markov Decision Process with discounted costs, and can be
solved using value iteration. In the finite horizon, we show that team performance is enhanced upon appropriate incentivization when compared to sequential social learning.
more »
« less
- Award ID(s):
- 1717530
- NSF-PAR ID:
- 10059998
- Date Published:
- Journal Name:
- Proceedings of the 2018 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
This work explores sequential Bayesian binary hypothesis testing in the social learning setup under expertise diversity. We consider a two-agent (say advisor-learner) sequential binary hypothesis test where the learner infers the hypothesis based on the decision of the advisor, a prior private signal, and individual belief. In addition, the agents have varying expertise, in terms of the noise variance in the private signal. Under such a setting, we first investigate the behavior of optimal agent beliefs and observe that the nature of optimal agents could be inverted depending on expertise levels. We also discuss suboptimality of the Prelec reweighting function under diverse expertise. Next, we consider an advisor selection problem wherein the belief of the learner is fixed and the advisor is to be chosen for a given prior. We characterize the decision region for choosing such an advisor and argue that a learner with beliefs varying from the true prior often ends up selecting a suboptimal advisor.more » « less
-
more » « less
Abstract Reinforcement learning is a general technique that allows an agent to learn an optimal policy and interact with an environment in sequential decision-making problems. The goodness of a policy is measured by its value function starting from some initial state. The focus of this paper was to construct confidence intervals (CIs) for a policy’s value in infinite horizon settings where the number of decision points diverges to infinity. We propose to model the action-value state function (Q-function) associated with a policy based on series/sieve method to derive its confidence interval. When the target policy depends on the observed data as well, we propose a SequentiAl Value Evaluation (SAVE) method to recursively update the estimated policy and its value estimator. As long as either the number of trajectories or the number of decision points diverges to infinity, we show that the proposed CI achieves nominal coverage even in cases where the optimal policy is not unique. Simulation studies are conducted to back up our theoretical findings. We apply the proposed method to a dataset from mobile health studies and find that reinforcement learning algorithms could help improve patient’s health status. A Python implementation of the proposed procedure is available at https://github.com/shengzhang37/SAVE.
-
In this paper, we consider federated Q-learning, which aims to learn an optimal Q-function by periodically aggregating local Q-estimates trained on local data alone. Focusing on infinite-horizon tabular Markov decision processes, we provide sample complexity guarantees for both the synchronous and asynchronous variants of federated Q-learning. In both cases, our bounds exhibit a linear speedup with respect to the number of agents and sharper dependencies on other salient problem parameters. Moreover, existing approaches to federated Q-learning adopt an equally-weighted averaging of local Q-estimates, which can be highly sub-optimal in the asynchronous setting since the local trajectories can be highly heterogeneous due to different local behavior policies. Existing sample complexity scales inverse proportionally to the minimum entry of the stationary state-action occupancy distributions over all agents, requiring that every agent covers the entire state-action space. Instead, we propose a novel importance averaging algorithm, giving larger weights to more frequently visited state-action pairs. The improved sample complexity scales inverse proportionally to the minimum entry of the average stationary state-action occupancy distribution of all agents, thus only requiring the agents collectively cover the entire state-action space, unveiling the blessing of heterogeneity.more » « less
-
We investigate the problem of persistently monitoring a finite set of targets with internal states that evolve with linear stochastic dynamics using a finite set of mobile agents. We approach the problem from the infinite-horizon perspective, looking for periodic movement schedules for the agents. Under linear dynamics and some standard assumptions on the noise distribution, the optimal estimator is a Kalman- Bucy filter. It is shown that when the agents are constrained to move only over a line and they can see at most one target at a time, the optimal movement policy is such that the agent is always either moving with maximum speed or dwelling at a fixed position. Periodic trajectories of this form admit finite parameterization, and we show to compute a stochastic gradient estimate of the performance with respect to the parameters that define the trajectory using Infinitesimal Perturbation Analysis. A gradient-descent scheme is used to compute locally optimal parameters. This approach allows us to deal with a very long persistent monitoring horizon using a small number of parameters.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
