US 12,293,296 B2
Adaptive path planning method based on neutral networks trained by the evolutional algorithms
Yongduan Song, Chongqing (CN); Lihui Tan, Chongqing (CN); Lei Fang, Chongqing (CN); Shilei Tan, Chongqing (CN); and Shuai Wang, Chongqing (CN)
Assigned to Dibi (Chongqing) Intelligent Technology Research Institute Co., Ltd., Chongqing (CN); and Star (Chongqing) Intelligent Equipment Technology Research Institute Co., Ltd., Chongqing (CN)
Filed by Dibi (chongqing) Intelligent Technology Research Institute Co., Ltd., Chongqing (CN); and Star (Chongqing) Intelligent Equipment Technology Research Institute Co., Ltd., Chongqing (CN)
Filed on Oct. 18, 2021, as Appl. No. 17/503,743.
Claims priority of application No. 202110974360.1 (CN), filed on Aug. 24, 2021.
Prior Publication US 2023/0062408 A1, Mar. 2, 2023
Int. Cl. G06N 3/086 (2023.01); G06N 3/044 (2023.01); G06N 3/045 (2023.01)
CPC G06N 3/086 (2013.01) [G06N 3/044 (2023.01); G06N 3/045 (2023.01)] 8 Claims
OG exemplary drawing
 
1. A neutral network training method based on the evolutional algorithms, comprising the following steps:
S100: constructing N recurrent neutral networks with the same structure, wherein the recurrent neutral networks comprising: the neutral networks adopt the A-tier structure, the input tier has a total of B beurons, the output tier has two outputs, and the neutral networks have a total of C link weights;
providing with N mobile robots, wherein each mobile robot is installed with the following acquisition structures: D target sensors are installed on the head of the mobile robot, the same number of the sonic sensors, with a total of E sonic sensors, are respectively installed on both sides of the mobile robot, the D target sensors and the E sonic sensors respectively correspond to the B beurons in the input tier, D+E=B, the data acquired by the D target sensors and the data acquired by the E sonic sensors are input into the B beurons;
S200: optimizing the recurrent neutral networks in S100 by using the evolutional algorithms as follows:
S210: initializing N recurrent neutral networks and obtaining kth individual by using the real-number encoding of the C link weights of the tth recurrent neutral network, wherein the kth individual is used as the first-generation chromosome of the evolutional algorithms, that is, paternal chromosome, t=k=1, 2 . . . N; setting the data acquisition step size of the D target sensors and the E sonic sensors, setting the maximum number of evolutions Gmax, setting the fitness threshold S, setting the number of populations N;
S220: determining the start point and the target point for N mobile robots in a location coordinate system, inputting the data acquired by the D target sensors and the data acquired by the E sonic sensors in each mobile robot into the B beurons of the input tier of the recurrent neutral networks in S210 in, and outputting linear velocity and angular velocity of the mobile robot at each acquisition point from the recurrent neutral networks;
S230: calculating the fitnesses of N mobile robots by using an evaluation function;
S240: selecting the recurrent neutral networks corresponding to the mobile robots with the greatest fitness, from the N recurrent neutral networks, and duplicating the paternal chromosomes corresponding to the recurrent neutral networks as progeny chromosomes Z1;
selecting the recurrent neutral networks corresponding to the mobile robots with the fitness less than the fitness threshold S, from the N recurrent neutral networks, and discarding the paternal chromosomes corresponding to the recurrent neutral networks;
selecting the recurrent neutral networks corresponding to the mobile robots with the fitness greater than or equal to the fitness threshold S, from the N recurrent neutral networks, duplicating one part of the paternal chromosomes corresponding to the recurrent neutral networks as the progeny chromosomes Z2, and obtaining the progeny chromosomes Z3 by dividing the other part of the paternal chromosomes into Part1 and Part2 and performing the evolutional operation of crossover and mutation respectively;
S250: judging whether the current number of evolutions is greater than the maximum number of evolutions Gmax; if yes, then execute the next step; if not, then use the progeny chromosomes Z1, the progeny chromosomes Z2 and the progeny chromosomes Z3 as the new paternal chromosomes and return to step S220;
S300: selecting the paternal chromosome corresponding to the maximum fitness value in each evolution to Gmax paternal chromosomes, and selecting a paternal chromosome with the maximum fitness value out of the Gmax paternal chromosomes as the global optimal individual;
S400: obtaining a global optimal neutral network based on the neutral network corresponding to the global optimal individual in the S300.