Vmware, Inc.

Gregory T Burk

Dev A Nag

Dongni Wang

Yanislav Yankov

Nicholas Mark Grant Stephen

Lachlan T Coote

Brian Alfred Williams

Alain M Dumesny

Yanislov Yankov

Michael Donahue

Janislav Jankov

Nick Stephen

The current document is directed to methods and systems for simulation-based training of automated reinforcement-learning-based application managers. Simulators are generated from data collected from controlled computing environments controlled and may employ any of a variety of different machine-learning models to learn state-transition and reward models. The current disclosed methods and systems provide facilities for visualizing aspects of the models learned by a simulator and for initializing simulator models using domain information. In addition, the currently disclosed simulators employ weighted differences computed from simulator-generated and training-data state transitions for feedback to the machine-learning models to address various biases and deficiencies of commonly employed difference metrics in the context of training automated reinforcement-learning-based application managers.

Simulator-training for automated reinforcement-learning-based application-managers

The current document is directed to an automated reinforcement-learning-based application manager that uses action tags and metric tags. In various implementations, actions and metrics are associated with tags. Different types of tags can contain different types of information that can be used to greatly improve the computational efficiency by which the reinforcement-learning-based application manager explores the action-state space in order to determine and maintain an optimal or near-optimal management policy by providing a vehicle for domain knowledge to influence control-policy decision making.

Automated reinforcement-learning-based application manager that uses action tags and metric tags

The current document is directed to a safe-operation-constrained reinforcement-learning-based application manager that can be deployed in various different computational environments, without extensive manual modification and interface development, to manage the computational environments with respect to one or more reward-specified goals. Control actions undertaken by the safe-operation-constrained reinforcement-learning-based application manager are constrained, by stored action filters, to constrain state/action-space exploration by the safe-operation-constrained reinforcement-learning-based application manager to safe actions and thus prevent deleterious impact to the managed computational environment.

Safe-operation-constrained reinforcement-learning-based application manager

The current document is directed to transfer of training received by a first automated reinforcement-learning-based application manager while controlling a first application is transferred to a second automated reinforcement-learning-based application manager which controls a second application different from the first application. Transferable training provides a basis for automated generation of applications from application components. Transferable training is obtained from composition of applications from application components and composition of reinforcement-learning-based-control-and-learning constructs from reinforcement-learning-based-control-and-learning constructs of application components.

Transferable training for automated reinforcement-learning-based application-managers

The current document is directed to automated reinforcement-learning-based application managers that that are trained using adversarial training. During adversarial training, potentially disadvantageous next actions are selected for issuance by an automated reinforcement-learning-based application manager at a lower frequency than selection of next actions, according to a policy that is learned to provide optimal or near-optimal control over a computing environment that includes one or more applications controlled by the automated reinforcement-learning-based application manager. By selecting disadvantageous actions, the automated reinforcement-learning-based application manager is forced to explore a much larger subset of the system-state space during training, so that, upon completion of training, the automated reinforcement-learning-based application manager has learned a more robust and complete optimal or near-optimal control policy than had the automated reinforcement-learning-based application manager been trained by simulators or using management actions and computing-environment responses recorded during previous controlled operation of a computing-environment.

Adversarial automated reinforcement-learning-based application-manager training

The current document is directed to automated reinforcement-learning-based application managers that use local agents. Local agents provide finer-granularity monitoring of an application or application subcomponents and provide continued application management in the event of interruption of network traffic between an automated reinforcement-learning-based application manager and the application or application subcomponents managed by the automated reinforcement-learning-based application manager.

Automated reinforcement-learning-based application manager that uses local agents

The current document is directed to automated reinforcement-learning-based application managers that learn and improve the reward function that steers reinforcement-learning-based systems towards optimal or near-optimal policies. Initially, when the automated reinforcement-learning-based application manager is first installed and launched, the automated reinforcement-learning-based application manager may rely on human-application-manager action inputs and resulting state/action trajectories to accumulate sufficient information to generate an initial reward function. During subsequent operation, when it is determined that the automated reinforcement-learning-based application manager is no longer following a policy consistent with the type of management desired by human application managers, the automated reinforcement-learning-based application manager may use accumulated trajectories to improve the reward function.

Automated reinforcement-learning-based application manager that learns and improves a reward function

The current document is directed to automated reinforcement-learning-based application managers that obtain increased computational efficiency by reusing learned models and by using human-management experience to truncate state and observation vectors. Learned models of managed environments that receive component-associated inputs can be partially or completely reused for similar environments. Human managers and administrators generally use only a subset of the available metrics in managing an application, and that subset can be used as an initial subset of metrics for learning an optimal or near-optimal control policy by an automated reinforcement-learning-based application manager.

Computationally efficient reinforcement-learning-based application manager

The current document is directed to an administrator-monitored reinforcement-learning-based application manager that can be deployed in various different computational environments to manage the computational environments with respect to one or more reward-specified goals. Certain control actions undertaken by the administrator-monitored reinforcement-learning-based application manager are first proposed, to one or more administrators or other users, who can accept or reject the proposed control actions prior to their execution. The reinforcement-learning-based application manager can therefore continue to explore the state/action space, but the exploration can be parametrically constrained as well as by human-administrator oversight and intervention.

Administrator-monitored reinforcement-learning-based application manager

The current document is directed to a modular reinforcement-learning-based application manager that can be deployed in various different computational environments without extensive modification and interface development. The currently disclosed modular reinforcement-learning-based application manager interfaces to observation and action adapters and metadata that provide a uniform and, in certain implementations, self-describing external interface to the various different computational environments which the modular reinforcement-learning-based application manager may be operated to control. In addition, certain implementations of the currently disclosed modular reinforcement-learning-based application manager interface to a user-specifiable reward-generation interface to allow the rewards that provide feedback from the computational environment to the modular reinforcement-learning-based application manager to be tailored to meet a variety of different user expectations and desired control policies.

Modular reinforcement-learning-based application manager

Some embodiments provide method for managing a set of computing resources. The method receives information for a set of resources. The information for each resource indicates a set of policies bound to the resource. The policies as bound to the resources are for application by several policy engines. For each of several of the resources, the method determines whether the policies bound to the resource violate a set of policy validation rules. For a subset of the resources for which a violation exists, the method disables at least one of the policies from being applied to the resource by the several policy engines.

Policy validation

Some embodiments provide, for a policy framework that manages application of a plurality of policies to a plurality of resources in a computing environment, a method for providing a user interface. The method displays a first display area for viewing and editing policies imported by the policy framework from a first several heterogeneous sources. The method displays a second display area for viewing and editing information regarding computing resources imported by the policy framework from a second several heterogeneous sources. The method displays a third display area for viewing and editing binding rules for binding the policies to the computing resources.

Policy framework user interface

Some embodiments provide a method for managing a set of computing resources. The method imports descriptions of computing resources from several heterogeneous sources. The descriptions specify categories for the computing resources. The different sources use different types of categories for the resource descriptions. Based on the categories specified for the computing resources in the set, the method derives additional categories for at least a subset of the computing resources. The method stores each imported computing resource tagged according to its specified and derived categories, wherein the category tags are used for binding policies to the computing resources.

Resource categorization for policy framework

Some embodiments provide a method for managing policies for a set of computing resources. The method imports several sets of resource management policy rules from several heterogeneous sources. The method stores each set of imported policy rules as a primitive policy. The primitive policies are (i) applicable to resources in the set of computing resources and (ii) combinable into composite policies that are applicable to resources in the set of computing resources. Composite policies are combinable into additional composite policies with primitive policies and other composite policies.

Policy store

Some embodiments provide, for a policy framework, a method for managing policies for a set of resources in a computing environment. The method stores several imported policy rules as primitive policies, each of which includes a policy data structure that includes a set of fields. One of the fields of each primitive policy stores the imported policy rule for the primitive policy. The method defines several composite policies based at least in part on the primitive policies. The method stores the defined composite policies as policy data structures. Each policy data structure for a composite policy includes a set of fields and references at least one additional policy data structure.

Growing community of inventors

Colorado Springs, CO, United States of America

Gregory T Burk