Client-side video players employ adaptive bitrate (ABR) algorithms to optimize user quality of experience (QoE). We evaluate recently proposed RL-based ABR methods in Facebook's web-based video streaming platform. Real-world ABR contains several challenges that requires customized designs beyond off-the-shelf RL algorithms -- we implement a scalable neural network architecture that supports videos with arbitrary bitrate encodings; we design a training method to cope with the variance resulting from the stochasticity in network conditions; and we leverage constrained Bayesian optimization for reward shaping in order to optimize the conflicting QoE objectives. In a week-long worldwide deployment with more than 30 million video streaming sessions, our RL approach outperforms the existing human-engineered ABR algorithms.

Real-time video inference on compute-limited edge devices like mobile phones and drones is challenging due to the high computation cost of Deep Neural Network models. In this paper we propose Adaptive Model Streaming (AMS), a cloud-assisted approach to real-time video inference on edge devices. The key idea in AMS is to use online learning to continually adapt a lightweight model running on an edge device to boost its performance on the video scenes in real-time. The model is trained in a cloud server and is periodically sent to the edge device. We discuss the challenges of online learning for video and present a practical design that takes into account the edge device, cloud server, and network bandwidth resource limitations. On the task of video semantic segmentation, our experimental results show 5.1--17.0 percent mean Intersection-over-Union improvement compared to a pre-trained model on several real-world videos. Our prototype can perform video segmentation at 30 frames-per-second with 40 milliseconds camera-to-label latency on a Samsung Galaxy S10+ mobile phone, using less than 400Kbps uplink and downlink bandwidth on the device.

We present Placeto, a reinforcement learning (RL) approach to efficiently find device placements for distributed neural network training. Unlike prior approaches that only find a device placement for a specific computation graph, Placeto can learn generalizable device placement policies that can be applied to any graph. We propose two key ideas in our approach: (1) we represent the policy as performing iterative placement improvements, rather than outputting a placement in one shot; (2) we use graph embeddings to capture relevant information about the structure of the computation graph, without relying on node labels for indexing. These ideas allow Placeto to train efficiently and generalize to unseen graphs. Our experiments show that Placeto requires up to 6.1x fewer training steps to find placements that are on par with or better than the best placements found by prior approaches. Moreover, Placeto is able to learn a generalizable placement policy for any given family of graphs, which can then be used without any retraining to predict optimized placements for unseen graphs from the same family. This eliminates the large overhead incurred by prior RL approaches whose lack of generalizability necessitates re-training from scratch every time a new graph is to be placed.

Symbol detection for Massive Multiple-Input Multiple-Output (MIMO) is a challenging problem for which traditional algorithms are either impractical or suffer from performance limitations. Several recently proposed learning-based approaches achieve promising results on simple channel models (e.g., i.i.d. Gaussian). However, their performance degrades significantly on real-world channels with spatial correlation. We propose MMNet, a deep learning MIMO detection scheme that significantly outperforms existing approaches on realistic channels with the same or lower computational complexity. MMNet's design builds on the theory of iterative soft-thresholding algorithms and uses a novel training algorithm that leverages temporal and spectral correlation to accelerate training. Together, these innovations allow MMNet to train online for every realization of the channel. On i.i.d. Gaussian channels, MMNet requires two orders of magnitude fewer operations than existing deep learning schemes but achieves near-optimal performance. On spatially-correlated channels, it achieves the same error rate as the next-best learning scheme (OAMPNet) at 2.5dB lower SNR and with at least 10x less computational complexity. MMNet is also 4--8dB better overall than a classic linear scheme like the minimum mean square error (MMSE) detector.

Efficiently scheduling data processing jobs on distributed compute clusters requires complex algorithms. Current systems, however, use simple generalized heuristics and ignore workload structure, since developing and tuning a bespoke heuristic for each workload is infeasible. In this paper, we show that modern machine learning techniques can generate highly-efficient policies automatically. Decima uses reinforcement learning (RL) and neural networks to learn workload-specific scheduling algorithms without any human instruction beyond specifying a high-level objective such as minimizing average job completion time. Off-the-shelf RL techniques, however, cannot handle the complexity and scale of the scheduling problem. To build Decima, we had to develop new representations for jobs' dependency graphs, design scalable RL models, and invent new RL training methods for continuous job arrivals. Our prototype integration with Spark on a 25-node cluster shows that Decima outperforms several heuristics, including hand-tuned ones, by at least 21%. Further experiments with an industrial production workload trace demonstrate that Decima delivers up to a 17% reduction in average job completion time and scales to large clusters.

We consider reinforcement learning in input-driven environments, where an exogenous, stochastic input process affects the dynamics of the system. Input processes arise in many applications, including queuing systems, robotics control with disturbances, and object tracking. Since the state dynamics and rewards depend on the input process, the state alone provides limited information for the expected future returns. Therefore, policy gradient methods with standard state-dependent baselines suffer high variance during training. We derive a bias-free, input-dependent baseline to reduce this variance, and analytically show its benefits over state-dependent baselines. We then propose a meta-learning approach to overcome the complexity of learning a baseline that depends on a long sequence of inputs. Our experimental results show that across environments from queuing systems, computer networks, and MuJoCo robotic locomotion, input-dependent baselines consistently improve training stability and result in better eventual policies.