MDDE is developed to facilitate the application of the reinforcement learning (RL) algorithms for the optimization of data distribution within a distributed data storage. The project is designed to represent a real-world data storage system instead of relying on cost models, which might not always be accurate.
The current implementation of MDDE focuses on read-oriented scenarios.
./registry
Since the goal is the creation of an environment for assessment and design of reinforcement learning algorithms, abstraction from existing database engines and the extensive heuristic optimization and distribution algorithms, which are built-in most of these, was one of the priorities. Additionally, an RL algorithm must have fine-grained control over the distribution and replication of data records. In order o satisfy these requirements, a simple distributed data orchestration module is provided: Registry.
./mdde
One of the primary goals in development was the simplification of integration with RL frameworks or algorithm implementation, while at the same time maintaining a high level of extensibility and accessibility for the researchers in the field. Therefore, the stage module of MDDE is written in Python and responsible for forming fragments, processing, and transforming action and observation spaces. Stage instance communicates with the registry via TCP connection and can be deployed separately from it and data nodes.
In order to set up a local development and debug version of MDDE, please switch to the ./debug folder and follow the instructions in the README.md file there.
For examples of Docker-based experiments configurations using our provided sample code from ./mdde/samples, please proceed to ./docker folder.
We provide a wrapper around Actor-Attention-Critic for Multi-Agent Reinforcement Learning. You can find the plugin and code samples in the repo: mdde-MAAC.
Any additional samples and supporting utilities, such as a processor of the history of the observation, are published in the mdde-toolbox repository.