Welcome to MARS’s documentation!¶
Note
The MARS-Gym is in ‘beta’ and currently under active development. Improvements to the code or documentation are welcome!
MARS-Gym (MArketplace Recommender Systems Gym), a benchmark framework for modeling, training, and evaluating RL-based recommender systems for marketplaces.
Three main components composes the framework:
Data Engineering Module: A highly customizable module where the consumer can ingest and process a massive amount of data for learning using spark jobs.
Simulation Module: Holds an extensible module built on top of PyTorch to design learning architectures. It also possesses an OpenAI’s Gym environment that ingests the processed dataset to run a multi-agent system that simulates the targeted marketplace.
Evaluation Module: Provides a set of distinct perspectives on the agent’s performance. It presents traditional recommendation metrics, off-policy evaluation metrics, and fairness indicators. This component is powered by a user-friendly interface to facilitate the analysis and comparison betweenagents
