Containerization test method and system for reinforcement learning model

Abstract

The invention discloses a containerized testing method and system for a reinforcement learning model. The method comprises the following steps that 1) a test party makes a test environment Docker mirror image, a proxy Docker mirror image and a corresponding connection module, a callback module and an evaluation module according to a to-be-established test task, and then packages the connection module into a proxy Docker mirror image file template; 2) the test party creates a test task on the test platform and uploads a mirror image file to a test party server; 3) the tested party downloads a mirror image file training agent of the test task, integrates the trained agent to an agent Docker mirror image and uploads the Docker mirror image to a server of the test party; 4) the test party server adds or replaces the callback module and the evaluation module in the newly uploaded proxy Docker mirror image file, repackages the newly uploaded proxy Docker mirror image file to obtain a new proxy Docker mirror image, and starts to run the test task; and 5) the test party server transmits the test process data back to the test platform.

Description

technical field [0001] The invention belongs to the technical field of computer software, and in particular relates to a containerized testing method and system for reinforcement learning models. Background technique [0002] Reinforcement learning is a learning method that is closer to life in reality. Unlike "deep learning" technology, it does not use pre-labeled data, but guides behavior through the rewards obtained by the agent interacting with the environment. The goal is to make the agent Get as many rewards from the environment as possible to learn the optimal policy. The testing of reinforcement learning needs to rely on the environment of reinforcement learning. In principle, the difference between reinforcement learning and deep learning is that the former needs to interact with the test environment online in real time, generate corresponding behaviors based on environmental feedback, and then make relevant judgments and evaluations; while the latter is not depend...

AI Technical Summary

Problems solved by technology

[0003] The traditional reinforcement learning model testing needs to build a test environment in the test server and connect the test objects to the test server. The problem with this approach is: First, from the requirements of the test environment, the testing process of the reinforcement learning model relies heavily on The test environment and its configuration files generally have high requirements for the test environment; secondly, in terms of the isolation of the test, the traditional test tightly binds the test object to the test software and hardware environment, and the isolation of the test process Poor, once the test environment needs to be replaced during the test process, it will cause repeated workload to build the test environment

[0004] At present, in the testing of traditional reinforcement learning models, the tested reinforcement learning model is heavily dependent on the test environment and the test fairness caused by software and hardware environment dependencies.

Images