A Cramér Distance perspective on Quantile Regression based Distributional Reinforcement Learning
Authors:
Alix Lhéritier,
Nicolas Bondoux
Abstract:
Distributional reinforcement learning (DRL) extends the value-based approach by approximating the full distribution over future returns instead of the mean only, providing a richer signal that leads to improved performances. Quantile Regression (QR) based methods like QR-DQN project arbitrary distributions into a parametric subset of staircase distributions by minimizing the 1-Wasserstein distance…
▽ More
Distributional reinforcement learning (DRL) extends the value-based approach by approximating the full distribution over future returns instead of the mean only, providing a richer signal that leads to improved performances. Quantile Regression (QR) based methods like QR-DQN project arbitrary distributions into a parametric subset of staircase distributions by minimizing the 1-Wasserstein distance. However, due to biases in the gradients, the quantile regression loss is used instead for training, guaranteeing the same minimizer and enjoying unbiased gradients. Non-crossing constraints on the quantiles have been shown to improve the performance of QR-DQN for uncertainty-based exploration strategies. The contribution of this work is in the setting of fixed quantile levels and is twofold. First, we prove that the Cramér distance yields a projection that coincides with the 1-Wasserstein one and that, under non-crossing constraints, the squared Cramér and the quantile regression losses yield collinear gradients, shedding light on the connection between these important elements of DRL. Second, we propose a low complexity algorithm to compute the Cramér distance.
△ Less
Submitted 22 February, 2022; v1 submitted 1 October, 2021;
originally announced October 2021.
From Research to Proof-of-Concept: Analysis of a Deployment of FPGAs on a Commercial Search Engine
Authors:
Fabio Maschi,
Gustavo Alonso,
Anthony Hock-Koon,
Nicolas Bondoux,
Teddy Roy,
Mourad Boudia,
Matteo Casalino
Abstract:
FPGAs are quickly becoming available in the cloud as a one more heterogeneous processing element complementing CPUs and GPUs. There are many reports in the literature showing the potential for FPGAs to accelerate a wide variety of algorithms, which combined with their growing availability, would seem to also indicate a widespread use in many applications. Unfortunately, there is not much published…
▽ More
FPGAs are quickly becoming available in the cloud as a one more heterogeneous processing element complementing CPUs and GPUs. There are many reports in the literature showing the potential for FPGAs to accelerate a wide variety of algorithms, which combined with their growing availability, would seem to also indicate a widespread use in many applications. Unfortunately, there is not much published research exploring what it takes to integrate an FPGA into an existing application in a cost-effective way and keeping the algorithmic performance advantages. Building on recent results exploring how to employ FPGAs to improve the search engines used in the travel industry, this paper analyses the end-to-end performance of the search engine when using FPGAs, as well as the necessary changes to the software and the cost of such deployments. The results provide important insights on current FPGA deployments and what needs to be done to make FPGAs more widely used. For instance, the large potential performance gains provided by an FPGA are greatly diminished in practice if the application cannot submit request in the most optimal way, something that is not always possible and might require significant changes to the application. Similarly, some existing cloud deployments turn out to use a very imbalanced architecture: a powerful FPGA connected to a not so powerful CPU. The result is that the CPU cannot generate enough load for the FPGA, which potentially eliminates all performance gains and might even result in a more expensive system. In this paper, we report on an extensive study and development effort to incorporate FPGAs into a search engine and analyse the issues encountered and their practical impact. We expect that these results will inform the development and deployment of FPGAs in the future by providing important insights on the end-to-end integration of FPGAs within existing systems.
△ Less
Submitted 20 August, 2021;
originally announced August 2021.