Model Selection for Off-policy Evaluation: New Algorithms and Experimental Protocol

Holdout validation and hyperparameter tuning from data is a long-standing problem in offline reinforcement learning (RL). A standard framework is to use off-policy evaluation (OPE) methods to evaluate and select the policies, but OPE either incurs exponential variance (e.g., importance sampling) or has hyperparameters on their own (e.g., FQE and model-based). In this work we focus on hyperparameter tuning for OPE itself, which is even more under-investigated. Concretely, we select among candidate value functions ("model-free") or dynamics ("model-based") to best assess the performance of a target policy. Our contributions are two fold. We develop: (1) new model-free and model-based selectors with theoretical guarantees, and (2) a new experimental protocol for empirically evaluating them. Compared to the model-free protocol in prior works, our new protocol allows for more stable generation of candidate value functions, better control of misspecification, and evaluation of model-free and model-based methods alike. We exemplify the protocol on a Gym environment, and find that our new model-free selector, LSTD-Tournament, demonstrates promising empirical performance.

Learning Estimates At The Edge Using Intermittent And Aged Measurement Updates

Cyber Physical Systems (CPS) applications have agents that actuate in their local vicinity, while requiring measurements that capture the state of their larger environment to make actuation choices. These measurements are made by sensors and communicated over a network as update packets. Network resource constraints dictate that updates arrive at an agent intermittently and be aged on their arrival. This can be alleviated by providing an agent with a fast enough rate of estimates of the measurements. Often works on estimation assume knowledge of the dynamic model of the system being measured. However, as CPS applications become pervasive, such information may not be available in practice. In this work, we propose a novel deep neural network architecture that leverages Long Short Term Memory (LSTM) networks to learn estimates in a model-free setting using only updates received over the network. We detail an online algorithm that enables training of our architecture. The architecture is shown to provide good estimates of measurements of both a linear and a non-linear dynamic system. It learns good estimates even when the learning proceeds over a generic network setting in which the distributions that govern the rate and age of received measurements may change significantly over time. We demonstrate the efficacy of the architecture by comparing it with the baselines of the Time-varying Kalman Filter and the Unscented Kalman Filter. The architecture enables empirical insights with regards to maintaining the ages of updates at the estimator, which are used by it and also the baselines.