Larry Rudolph

Logan Engstrom, Andrew Ilyas, Shibani Santurkar et al.

We study the roots of algorithmic progress in deep policy gradient algorithms through a case study on two popular algorithms: Proximal Policy Optimization (PPO) and Trust Region Policy Optimization (TRPO). Specifically, we investigate the consequences of "code-level optimizations:" algorithm augmentations found only in implementations or described as auxiliary details to the core algorithm. Seemingly of secondary importance, such optimizations turn out to have a major impact on agent behavior. Our results show that they (a) are responsible for most of PPO's gain in cumulative reward over TRPO, and (b) fundamentally change how RL methods function. These insights show the difficulty and importance of attributing performance gains in deep reinforcement learning. Code for reproducing our results is available at https://github.com/MadryLab/implementation-matters .

Amin Mosayyebzadeh, Gerardo Ravago, Apoorve Mohan et al.

Bolted is a new architecture for a bare metal cloud with the goal of providing security-sensitive customers of a cloud the same level of security and control that they can obtain in their own private data centers. It allows tenants to elastically allocate secure resources within a cloud while being protected from other previous, current, and future tenants of the cloud. The provisioning of a new server to a tenant isolates a bare metal server, only allowing it to communicate with other tenant's servers once its critical firmware and software have been attested to the tenant. Tenants, rather than the provider, control the tradeoffs between security, price, and performance. A prototype demonstrates scalable end-to-end security with small overhead compared to a less secure alternative.

Amin Mosayyebzadeh, Apoorve Mohan, Sahil Tikale et al.

Bolted is a new architecture for bare-metal clouds that enables tenants to control tradeoffs between security, price, and performance. Security-sensitive tenants can minimize their trust in the public cloud provider and achieve similar levels of security and control that they can obtain in their own private data centers. At the same time, Bolted neither imposes overhead on tenants that are security insensitive nor compromises the flexibility or operational efficiency of the provider. Our prototype exploits a novel provisioning system and specialized firmware to enable elasticity similar to virtualized clouds. Experimentally we quantify the cost of different levels of security for a variety of workloads and demonstrate the value of giving control to the tenant.

Andrew Ilyas, Logan Engstrom, Shibani Santurkar et al.

We study how the behavior of deep policy gradient algorithms reflects the conceptual framework motivating their development. To this end, we propose a fine-grained analysis of state-of-the-art methods based on key elements of this framework: gradient estimation, value prediction, and optimization landscapes. Our results show that the behavior of deep policy gradient algorithms often deviates from what their motivating framework would predict: the surrogate objective does not match the true reward landscape, learned value estimators fail to fit the true value function, and gradient estimates poorly correlate with the "true" gradient. The mismatch between predicted and empirical behavior we uncover highlights our poor understanding of current methods, and indicates the need to move beyond current benchmark-centric evaluation methods.