Robot optimal path planning method based on partially observable Markov decision process

Abstract

The invention discloses a robot optimal path planning method based on a partially observable Markov decision process. A robot searches an optimal path to a target position, a POMDP model and a SARSOPalgorithm are considered as a basis, and a GLS search method is utilized as a heuristic condition during searching. For a continuous state large-scale observation space problem, through usage of the robot optimal path planning method the number of times of belief upper and lower bound updating is reduced than that of early classical algorithms adopting test as a heuristic condition to repeat updating of multiple similar paths, and final optimal strategy is not affected, thereby improving the algorithm efficiency; and in the same time, the robot can get a better strategy and find a better path.

Description

technical field [0001] The invention relates to the field of robot control, in particular to a robot optimal path planning method based on a partially perceptual Markov decision process. Background technique [0002] Machine Learning (ML) is a discipline that studies how to simulate or realize human learning behavior, and constantly reorganize and improve its original knowledge structure. Reinforcement learning is an important research branch of machine learning. It is a machine learning method that maps state to action through the interaction between agent and environment, so as to obtain the maximum long-term cumulative discount reward. Usually reinforcement learning uses Markov decision processes (Markov Decision Processes, MDPs) as a model, that is, the environment is completely observable. In the real world, however, uncertainty is ubiquitous. For example, the agent's sensor has its own limitations: (1) the sensor can only detect local limited environments, and the ag...

AI Technical Summary

Problems solved by technology

However, the trial-based search selects the optimal action and observation each time, without considering other observations that are very close to the optimal observation and have a significant impact on future algorithm performance

In large-scale observation space problems, the performance of the algorithm is poor

Images