Google AI Introduces ArchGym: An Open-Source Gymnasium for Machine Learning that Connects a Diverse Range of Search Algorithms To Architecture Simulators

Abstraction: Open-source ML gymnasium for reproducible computer architecture design space exploration

Key points:

• ArchGym provides a unified, open-source interface connecting ML search algorithms (random walk, RL, Bayesian optimization) to architecture simulators (DRAM controllers, DNN accelerators, etc.)
• Two components: environment (encapsulates architecture cost model and workloads) and agent (holds hyperparameters and policy); communicate via hardware status, parameters, and metrics signals
• Key finding: at least one hyperparameter configuration of any search algorithm can match the best-performing ML method; wrong algorithm conclusions arise from poor hyperparameter choices
• Addresses reproducibility gap in architecture research: prior work lacked robust baselines enabling fair cross-methodology comparison
• ArchGym Dataset stores all design space exploration outputs for artifact generation and analysis
• Available at github.com/srivatsankrishnan/oss-arch-gym; published at ISCA 2023 (ACM DL DOI 10.1145/3579371.3589049)

Connections: Google · Neural Architecture Search · Design Space Exploration · Reinforcement Learning

Source: https://www.marktechpost.com/2023/07/14/google-ai-introduces-archgym-an-open-source-gymnasium-for-machine-learning-that-connects-a-diverse-range-of-search-algorithms-to-architecture-simulators/