The invention discloses a
Ray-based lightweight distributed
reinforcement learning training platform design method. The method comprises the steps of obtaining an
original data stream set, constructing topological feature representation, performing dynamic
smoothing processing, and generating stable feature representation. Calculating an information
retention rate and a topology retention degree, constructing a state manifold, a tangent space and a
priority queue, and updating an experience
pool; strategy parameters and gradients are collected, a Karcher mean value is calculated,
momentum correction is carried out, a parameter manifold
layered structure and interlayer mapping are constructed, and a parameter structure is optimized; constructing a
knowledge graph, and performing structured
pruning and quantitative optimization to obtain final compression
model parameters; constructing a
load distribution matrix, generating an initial
parameter distribution scheme, carrying out iterative optimization, constructing a load balancing objective function, generating a task scheduling strategy, and finally deploying and configuring. According to the method, the complexity of
data processing is effectively reduced, the calculation efficiency is improved, and the requirement for manual adjustment of developers is reduced through various automatic optimizations.