CN113514254A - Parallel acceleration test method for automatic driving simulation - Google Patents
Parallel acceleration test method for automatic driving simulation Download PDFInfo
- Publication number
- CN113514254A CN113514254A CN202110478572.0A CN202110478572A CN113514254A CN 113514254 A CN113514254 A CN 113514254A CN 202110478572 A CN202110478572 A CN 202110478572A CN 113514254 A CN113514254 A CN 113514254A
- Authority
- CN
- China
- Prior art keywords
- test
- iteration
- layer
- area
- parameter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M17/00—Testing of vehicles
- G01M17/007—Wheeled or endless-tracked vehicles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to a parallel acceleration test method aiming at automatic driving simulation. The method comprises the following steps: determining a parallel accelerated test architecture, dividing test areas and determining tasks of each level; step two, allocating the execution quantity of the bottom layers of different small areas to ensure the test effect of the whole parameter space; step three, determining the bottom performer test parameters in the small area to ensure the test effect in the small area; step four, defining the bottom layer execution generation corresponding to the test result; and step five, setting a test termination condition and interrupting the test process in time. The running capability of a plurality of computing platforms can be fully utilized by using a parallel testing means, and meanwhile, the generation of dangerous scenes can be further strengthened by optimizing algorithms in the whole area and the small area in the algorithm, so that the testing process of the algorithm to be tested in the logic scene is accelerated. The method can meet the requirements of the automatic driving test field on the parallel and accelerated test process, and has wide application prospect.
Description
Technical Field
The invention relates to the technical field of automobiles, in particular to a parallel acceleration test method aiming at automatic driving simulation.
Background
With the continuous maturity of the automatic driving technology, how to verify the safety of the automatic driving technology becomes a main obstacle restricting the mass production of the automatic driving automobile. Simulation tests based on scenes are a new idea for solving the problem of automatic driving tests, and are widely concerned by researchers.
The existing scene-based automatic driving simulation test method mostly focuses on construction of a simulation test process or single-line simulation accelerated test, and the accelerated test refers to actively searching scene parameter combinations which are possibly dangerous in a parameter space range so as to strengthen generation of dangerous scenes. With the increasing complexity of the automatic driving test scene, the spatial range of the scene parameters to be tested is increasing, which means that a large number of test cases are required to meet the confidence requirement of the test process. The current single-line test method cannot meet the requirement of large-scale test, and a parallel test method is urgently needed to apply the performance of the computing unit in a large scale so as to improve the test efficiency.
Disclosure of Invention
The invention provides a parallel acceleration test method aiming at automatic driving simulation, which solves the defect of the existing parallel acceleration test method.
The technical scheme of the invention is described as follows by combining the attached drawings:
a parallel acceleration test method for automatic driving simulation comprises the following steps:
determining a parallel accelerated test architecture, dividing test areas and determining tasks of each level;
step two, allocating the execution quantity of the bottom layers of different small areas to ensure the test effect of the whole parameter space;
step three, determining the bottom performer test parameters in the small area to ensure the test effect in the small area;
step four, defining the bottom layer execution generation corresponding to the test result;
and step five, setting a test termination condition and interrupting the test process in time.
The specific method of the first step is as follows:
in the test process, a tested logic scene is regarded as a complete large area, and the large area of the logic scene parameter space is uniformly divided into a certain number of small areas; establishing a parallel accelerated test dangerous scene by designing a three-layer architecture; the three-layer architecture is respectively a top-layer manager, a secondary manager and a bottom-layer executor; the top manager is a scheduling unit at the top layer of the algorithm and is responsible for mutual scheduling among bottom executors which are responsible for different secondary managers; the secondary manager is a middle-layer scheduling unit in the algorithm and is responsible for planning the next working content of different bottom-layer executors in the small area, namely the parameters of the next test, and meanwhile, the secondary manager can mark the easily lost person in the responsible area range; the easily lost person is easy to leave a small area which is responsible for the current secondary manager due to low output, and is allocated to a bottom executive of other small areas by a top manager, and the output is the danger level of the generated scene; the bottom executor is a calculation unit for executing the test specifically, and receives the test parameter requirement sent by the secondary manager and performs the test.
The specific method of the second step is as follows:
21) setting a top executive responsible for scheduling the bottom executive among different small areas, wherein the specific process comprises the following steps: monitoring the average output of the current iteration in each small area, calculating the average output of the whole large area in the same round, and commanding secondary managers in the area to allocate the easily lost people in the small area to go to the area with larger average output or the area generating extreme points for the small area with the average output lower than the average output of the whole area; the probability that a person who is easy to lose leaves different small production areas is as follows:
in the formula, piThe probability that the easily lost person in the ith cell leaves the cell is shown;average value of all bottom-layer executors output for j iteration; gijAverage yield in the ith cell for the jth iteration; j is the number of iterations; t is a regulating parameter which represents the strength of the top-level executor, the larger the value is, the larger the strength of the top-level manager is, the easier the secondary manager can listen to the deployment of the top-level manager, and the value is not less than 10; k is a pre-designed iteration number; n is the bottom layer executor number expected to be allocated to each small area by the top layer manager, and the value is not less than the initially allocated bottom layer executor number; n isijThe number of bottom-layer executors of the ith small area in the jth iteration is counted; n isijmaxThe number of the remaining active extreme values of the ith small region in the jth iteration is set; if a plurality of loss-prone persons exist in a region, the leaving probability of the loss-prone persons in one iteration is calculated sequentially, namely the leaving probability of the loss-prone person with the minimum yield in the current iteration is calculated firstly, whether the loss-prone person leaves is determined, and then the loss-prone person with the second minimum yield is calculated;
22) when all bottom executors in a cell run off, a top executor needs to define the cell as a development-value-free zone, seal the cell and forbid other bottom executors from entering the cell;
23) probability o of wandering off from the Small production region to a different region in round j +1 if there is anyiThe inflow calculation process is sequential calculation through the formula (2), and after each easily lost person flows in, the first term at the upper right side of the equation is changed;
aj=(aini-aend)·(K-j)/K+aend (3)
bj=(bini-bend)·(K-j)/K+bend (4)
cj=1-aj-bj (5)
in the formula, ninInputting the number of small areas; n isj *Is n of inputinThe total number of bottom performers of the small area in the jth round; gj *Is n of inputinTotal production of small regions in round j; lj *Is n of inputinThe total number of the maximum extreme values generated by the small region in the jth turn; a isj,bj,cjIn order to adjust parameters, the weights of the three sub-items are adjusted, the sum of the weights is 1, and the influence weight of the number of newly generated extreme points is larger and larger as iteration progresses; n isijThe number of bottom-layer executors of the ith small area in the jth iteration is counted; gijAverage yield in the ith cell for the jth iteration; lijThe maximum extreme value number in the ith cell of the jth iteration is obtained; gjAverage value of all bottom-layer executors output for j iteration; a isiniIs an initial extreme value weight coefficient; a isendA weight coefficient for a termination extremum; k is a pre-designed iteration number; j is the current iteration number; biniIs an initial output weight coefficient; bendTerminating the yield weight coefficient;
each new churner inflow changes the number of underlying executions of a region, requiring recalculation of the probability of flowing into a different region.
The concrete method of the third step is as follows:
31) the secondary manager is responsible for recording the maximum yield position, the minimum yield position, all tested parameter positions, all yields and corresponding turn numbers of each bottom-layer executor in the small area where the secondary manager is located, which are known up to now; each round, the secondary manager defines a plurality of convergence areas centered on the known maximum extremum in the area, and the bottom performers within the convergence areas, which are not at the maximum extremum, move at the following speed:
ωj=(ωini-ωend)·(K-j)/K+ωend (7)
vi0=rand(0,vmax) (8)
in the formula, xijThe test parameter position of the ith bottom-layer performer in the area in the jth iteration is determined; v. ofijThe speed of the ith bottom-layer performer in the jth iteration is obtained; dijIs a distance xijThe parameter location of the nearest maximum extremum; h isijStopping the position of the maximum output of the jth iteration for the ith bottom-layer executor; t is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86;taking 2.05 as a learning coefficient; omega is an inertia coefficient; omegainiTaking 0.9 as an initial inertia coefficient; omegaendTo terminate the inertia coefficient, 0.4 is taken; omegajThe velocity weight coefficient in the iteration process of the previous round is obtained; k is a preset iteration number; j is the current iteration number; v. ofi0A speed parameter for a first iteration; v. ofmaxIs a set maximum speed;
32) the convergence zone will increase with the increase of the test rounds, and the formula of the convergence zone range is:
rj+1=(2-T)·rj (10)
in the formula, all s meeting the conditions form a convergence region; u is different dimensions of the parameter space; k is all dimensions of the parameter space; r isuIs the axial length of the convergence region in the u-th dimension; r isjThe axial length of a convergence region of the j-th iteration; r is1Is the initial convergence region axial length; y is the maximum value of the parameters of different dimensions of the parameter space; e is the minimum value of the parameters of different dimensions of the parameter space; m is the number of all bottom layer executions; r isj+1The axial length of a convergence region of the j +1 th iteration; t is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86;
33) for the bottom performers at the non-maximum extreme outside the convergence region, the performers move at the following speeds:
in the formula, T is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86; omegajThe velocity weight coefficient in the iteration process of the previous round is obtained; v. ofijThe speed of the ith bottom-layer performer in the jth iteration is obtained;taking 2.05 as a learning coefficient; h isijStopping the position of the maximum output of the jth iteration for the ith bottom-layer executor; x is the number ofijThe test parameter position of the ith bottom-layer performer in the area in the jth iteration is determined;
searching a nearest untested position near a current round working point or a parameter point obtained by calculation by a bottom-layer executor moving to a non-specified parameter position through the calculation for exploration; directly selecting the unexperienced parameter closest to the bottom performer at the maximum extreme value as the next experimental parameter;
34) the secondary manager marks the probability f that the i-th bottom performer in the small area becomes an vulnerable personiAs shown in equation (13); the first term of the addition in the formula represents the comparison of the bottom performer's yield and the overall bottom performer's yield, the second term of the addition represents the comparison of the bottom performer's yield and the bottom performer's yield in the area, w is a crowding factor that indicates the distribution of the area in which the test has been performed, if anyThe Euclidean distance between the available closest point and the calculated point is greater than 2 times of the Euclidean distance of the initial discrete step length, if w is 1, otherwise, the value is 0;
wherein A, B is a regulation parameter, the sum of which is 1, and both of them are taken to be 0.5;average output of all bottom performers within 5 iterations; gi' is the average yield of the ith bottom performer over 5 iterations in the small area;average yield, w, for the bottom performers over 5 iterations in the regionijThe crowding coefficient of the ith bottom-layer performer in the small area in the j iteration is obtained; when the iteration number is less than 5, the average value of the running iteration number is taken; in addition thereto, forFor example, whether a specific bottom performer in the region is concerned about, that is, whether a loss-prone person enters or not, or the whole region condition in 5 iterations is concerned about, that is, the average yield of the latest 5 iterations is divided by 5;
since the vulnerable person does not necessarily leave the working area immediately, it is still possible to continue working in the area for a while if a vulnerable person finds a maximum point in the working phase of the area or exceeds the maximum point in the iteration processAndthe mark of the easy-to-lose person is lost and becomes the common bottom-layer executor again;
to prevent the vulnerable from frequent flowing between different areas, it is specified that an underlying actor will not be marked as vulnerable when the number of iterations in the current area is less than 5.
The concrete method of the fourth step is as follows:
the bottom layer executor receives the parameter point position sent by the secondary manager, puts the parameter at the parameter point position into the simulation platform for simulation, records whether the test process is collided or not, if not, records the maximum longitudinal TTC in the whole test process-1I.e. time to collision, maximum transverse TTC-1The yield G of the bottom performer in this experiment was:
wherein collision is a collision signal, which represents the occurrence of collision; LTTC-1For maximum longitudinal TTC of test car in test process-1;HTTC-1For maximum lateral TTC of test vehicle in test process-1。
The concrete method of the step five is as follows:
and (4) setting a test ending mark, reaching the set iteration number K, or finding no new extreme value in the continuous q rounds after a certain iteration number, wherein q is more than or equal to 2, and ending the test.
The invention has the beneficial effects that: the invention establishes a parallel acceleration test method aiming at automatic driving simulation, can fully utilize the operation capability of a plurality of computing platforms by using a parallel test means, and simultaneously, the generation of dangerous scenes can be further strengthened by optimizing algorithms in an integral area and a small area in the algorithm, thereby accelerating the test flow of the automatic driving tested algorithm in a logic scene. The method can meet the requirements of the automatic driving test field on the parallel and accelerated test process, and has wide application prospect.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and for those skilled in the art, other related drawings can be obtained according to the drawings without inventive efforts.
FIG. 1 is a schematic structural view of a three-layer frame;
FIG. 2 is a flow chart of the determination of the input area and the output area of the vulnerable person;
FIG. 3 is a flow chart of the judgment of the person who is easy to lose;
FIG. 4 is a flow chart illustrating the determination of the flow of the lost people;
FIG. 5 is a flow chart of the calculation of the work position of the underlying performer;
FIG. 6 is a schematic structural diagram of a cut-in scene of a leading vehicle in the embodiment;
FIG. 7 is a graph showing the test results of this example.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
A parallel acceleration test method for automatic driving simulation comprises the following steps:
determining a parallel accelerated test architecture, dividing test areas and determining tasks of each level;
establishing a parallel accelerated test dangerous scene by designing a three-layer architecture; the three-layer architecture is respectively a top-layer manager, a secondary manager and a bottom-layer executor; the top manager is a scheduling unit at the top layer of the algorithm and is responsible for mutual scheduling among bottom executors which are responsible for different secondary managers; for example, the underlying actor 1 is scheduled from the zone for which secondary administrator a is responsible to the zone for which secondary administrator B is responsible. The secondary manager is a middle-layer scheduling unit in the algorithm and is responsible for planning the next working content of different bottom-layer executors in the small area, namely the parameters of the next test, for example, the current experiment parameter of the bottom-layer executor is M, the secondary manager can require the next experiment parameter of the bottom-layer executor to be N, and meanwhile, the secondary manager can mark the easily lost person in the responsible area range; the easily lost person is easy to leave a small area which is responsible for the current secondary manager due to low output, and is allocated to a bottom executive of other small areas by a top manager, and the output is the danger level of the generated scene; the bottom executor is a calculation unit for executing the test specifically, and receives the test parameter requirement sent by the secondary manager and performs the test. In the test process, a tested logic scene is regarded as a complete large area, and in order to prevent a bottom-layer executor from being excessively concentrated on a certain part in the early stage of a test and not finding a dangerous scene of other key positions, the large area of a logic scene parameter space is uniformly divided into a certain number of small areas according to a space range; therefore, each area in the previous period is guaranteed to have a certain number of bottom-layer executors to fully explore, meanwhile, in order to prevent the generated scene parameters from being too close to reduce the calculation efficiency, discretization processing is carried out on the parameter areas, and discrete step length is selected according to test requirements to obtain a discrete parameter set.
That is, the whole logic scene parameter space is as the whole large rectangle in the left side of fig. 1, the top-level executor decomposes the whole large rectangle into a plurality of small rectangles, and controls the scheduling of the bottom-level executor in different small rectangles; the secondary manager manages the next operation of different bottom managers in a small rectangle; the bottom performer is responsible for performing specific experiments in a small area.
Step two, allocating the execution quantity of the bottom layers of different small areas to ensure the test effect of the whole parameter space;
in order to ensure that a bottom-layer executor can move from a small area with less output to a small area with larger output, thereby strengthening the generation of a dangerous scene, a top-layer executor is set up to be responsible for the scheduling of the bottom-layer executor among different small areas, and the specific flow is as follows:
referring to fig. 2-4, the average yield of each cell in the current iteration is monitored, and the average yield of the entire large area is calculated, and for the small areas with average yield lower than the average yield of the whole area, the secondary managers in the area are instructed to allocate the vulnerable to loss in the small areas to go to the areas with larger average yield or the areas with extreme points.
The probability that a person who is easy to lose leaves different small production areas is as follows:
in the formula, piThe probability that the easily lost person in the ith cell leaves the cell is shown;average value of all bottom-layer executors output for j iteration; gijAverage yield in the ith cell for the jth iteration; j is the number of iterations; t is a regulating parameter which represents the strength of the top-level executor, the larger the value is, the larger the strength of the top-level manager is, the easier the secondary manager can listen to the deployment of the top-level manager, and the value is not less than 10; k is a pre-designed iteration number; n is the bottom layer executor number expected to be allocated to each small area by the top layer manager, and the value is not less than the initially allocated bottom layer executor number; n isijThe number of bottom-layer executors of the ith small area in the jth iteration is counted; n isijmaxThe number of remaining active extreme values in the j iteration for the ith small region.
If there are many effusive persons in a region, their leaving probabilities in an iteration are calculated sequentially, that is, the leaving probability of the effusive person with the smallest yield in the current iteration is calculated first, whether it leaves is determined, then the effusive person with the second smallest yield is calculated, and so on.
It should be noted that there may be many churners in a region, and their leaving probabilities in one iteration are calculated sequentially, i.e. the leaving probability of the churner with the smallest yield in the current iteration is calculated first, after determining whether it leaves, the second smallest churner with the smallest yield is calculated, and so on.
When all the bottom performers in a small area are lost, the top performer needs to define the area as a development-value-free area, seal the small area, and prohibit other bottom performers from entering the small area.
Probability o of wandering off from the Small production region to a different region in round j +1 if there is anyiAs shown in equation (2), the inflow calculation process is also a sequential calculation, and the first term at the upper right side of the equation changes after each vulnerable person flows in.
aj=(aini-aend)·(K-j)/K+aend (3)
bj=(bini-bend)·(K-j)/K+bend (4)
cj=1-aj-bj (5)
In the formula, ninInputting the number of small areas; n isj *Is n of inputinThe total number of bottom performers of the small area in the jth round; gj *Is n of inputinTotal production of small regions in round j; lj *Is n of inputinThe total number of the maximum extreme values generated by the small region in the jth turn; a isj,bj,cjIn order to adjust parameters, the weights of the three sub-items are adjusted, the sum of the weights is 1, and the influence weight of the number of newly generated extreme points is larger and larger as iteration progresses; n isijThe number of bottom-layer executors of the ith small area in the jth iteration is counted; gijAverage yield in the ith cell for the jth iteration; lijThe maximum extreme value number in the ith cell of the jth iteration is obtained;average value of all bottom-layer executors output for j iteration; a isiniIs an initial extreme value weight coefficient; a isendA weight coefficient for a termination extremum; k is a pre-designed iteration number; j is the current iteration number; biniIs an initial output weight coefficient; bendTo terminate the yield weight coefficient;
it is also noted that each new churner inflow changes the number of underlying executions of a region, requiring recalculation of the probability of flowing into a different region.
Step three, determining the bottom performer test parameters in the small area to ensure the test effect in the small area;
referring to fig. 5, in order to ensure that the bottom performer in a small area also has the ability to search for a dangerous scene, a secondary manager is set to plan the test parameters of the bottom performer in a small area, so as to ensure that the bottom performer in a small area can also perform the intensified generation of the dangerous scene, and the specific work flow is as follows: the secondary manager is responsible for recording the maximum yield position, the minimum yield position, all tested parameter positions, all yields and corresponding turn numbers of each bottom-layer executor in the small area where the secondary manager is located, which are known up to now. Each round, the secondary manager defines a plurality of convergence areas centered on the known maximum extremum in the area, and the bottom performers within the convergence areas, which are not at the maximum extremum, move at the following speed:
ωj=(ωini-ωend)·(K-j)/K+ωend (7)
vi0=rand(0,vmax) (8)
in the formula, xijThe test parameter position of the ith bottom-layer performer in the area in the jth iteration is determined; v. ofijFor the ith bottom performer in the jth roundThe speed of the iteration; dijIs a distance xijThe parameter location of the nearest maximum extremum; h isijStopping the position of the maximum output of the jth iteration for the ith bottom-layer executor; t is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86;taking 2.05 as a learning coefficient; omega is an inertia coefficient; omegainiTaking 0.9 as an initial inertia coefficient; omegaendTo terminate the inertia coefficient, 0.4 is taken; omegajThe velocity weight coefficient in the iteration process of the previous round is obtained; k is a preset iteration number; j is the current iteration number; v. ofi0A speed parameter for a first iteration; v. ofmaxIs a set maximum speed;
the convergence zone will increase with the increase of the test rounds, and the formula of the convergence zone range is:
rj+1=(2-T)·rj (10)
in the formula, all s meeting the conditions form a convergence region; u is different dimensions of the parameter space; k is all dimensions of the parameter space; r isuIs the axial length of the convergence region in the u-th dimension; r isjThe axial length of a convergence region of the j-th iteration; r is1Is the initial convergence region axial length; y is the maximum value of the parameters of different dimensions of the parameter space; e is the minimum value of the parameters of different dimensions of the parameter space; m is the number of all bottom layer executions; r isj+1The axial length of a convergence region of the j +1 th iteration; t is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86;
for the bottom performers at the non-maximum extreme outside the convergence region, the performers move at the following speeds:
in the formula, T is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86; omegajThe velocity weight coefficient in the iteration process of the previous round is obtained; v. ofijThe speed of the ith bottom-layer performer in the jth iteration is obtained;taking 2.05 as a learning coefficient; h isijStopping the position of the maximum output of the jth iteration for the ith bottom-layer executor; x is the number ofijThe test parameter position of the ith bottom-layer performer in the area in the jth iteration is determined;
searching a nearest untested position near a current round working point or a parameter point obtained by calculation by a bottom-layer executor moving to a non-specified parameter position through the calculation for exploration; and directly selecting the unexperienced parameter closest to the bottom-layer performer at the maximum extreme value as the next experimental parameter.
The secondary manager marks the probability f that the i-th bottom performer in the small area becomes an vulnerable personiAs shown in equation (13); the first term of addition in the formula represents the comparison between the output of the bottom performer and the output of all the bottom performers, the second term of addition represents the comparison between the output of the bottom performer and the output of the bottom performers in the area, w is a crowding coefficient and is used for representing the distribution condition of a tested area in the area, if the Euclidean distance between the available nearest point and the calculated point is greater than 2 times of the Euclidean distance of the initial discrete step length, w is 1, otherwise, the value is 0;
wherein A, B is a regulation parameter, the sum of which is 1, and both of them are taken to be 0.5;average output of all bottom performers within 5 iterations; gi' is the average yield of the ith bottom performer over 5 iterations in the small area;average yield, w, for the bottom performers over 5 iterations in the regionijThe crowding coefficient of the ith bottom-layer performer in the small area in the j iteration is obtained; when the iteration times are less than 5, the average value of the operated iteration times is obtained; in addition thereto, forFor example, whether a specific bottom-layer performer in the region enters the region is not concerned, and the overall region condition in 5 iterations is concerned, that is, the average yield of the last 5 iterations is divided by 5.
Since the vulnerable person does not necessarily leave the working area immediately, it is still possible to continue working in the area for a while if a vulnerable person finds a maximum point in the working phase of the area or exceeds the maximum point in the iteration processAndthe mark of the easy-to-lose person is lost and becomes the common bottom-layer executor again;
to prevent the vulnerable from frequent flowing between different areas, it is specified that an underlying actor will not be marked as vulnerable when the number of iterations in the current area is less than 5.
I.e., the specific workflow of the secondary administrator to manage the work location of the underlying performers is shown in fig. 5. Firstly, the position of the maximum extreme point in the responsible cell is judged, then a convergence region is divided by taking the position as the center, the bottom-layer executor in the convergence region moves along the direction step length in the formula (6), and the bottom-layer executor outside the convergence region moves along the direction step length obtained by calculation in the formula (12).
Step four, defining the bottom layer execution generation corresponding to the test result;
the bottom performer is a computing platform for carrying out specific tests, is provided with an automatic driving simulation test tool, comprises scene construction software, an algorithm execution platform and the like, determines the output of the parameter position according to the test result, and has the following specific working procedures:
the bottom layer executor receives the parameter point position sent by the secondary manager, puts the parameter at the parameter point position into the simulation platform for simulation, records whether the test process is collided or not, if not, records the maximum longitudinal TTC in the whole test process-1I.e. time to collision, maximum transverse TTC-1The yield G of the bottom performer in this experiment was:
wherein collision is a collision signal, which represents the occurrence of collision; LTTC-1For maximum longitudinal TTC of test car in test process-1;HTTC-1For maximum lateral TTC of test vehicle in test process-1。
Step five, the test is finished;
in order to ensure that the algorithm can stop iteration in time so as to save calculation consumption, the method establishes a test ending mark, and the test ending mark is specifically set as follows: and when the set iteration number K is reached, or after a certain iteration number, no new extreme value is found in the continuous q rounds, and q is more than or equal to 2.
Examples
As an example of the algorithm, a preceding vehicle cut-in scene is taken, and as shown in FIG. 6, the speed v of the vehicle after cut-in is selected1Distance d between cut-in front vehicle and speed v of cut-in front vehicle2As scene parameters, their parameter spaces are [14m/s, 38m/s, respectively]、[5m,55m]、[18.5m/s,45.5m/s]. Separation of velocityAnd selecting 3m/s for discrete step length and 1m for discrete step length of distance to obtain 4590 specific scene parameters. The number of selected bottom layer executions is 25. After 35 iterations, the experiment was automatically exited, taking 1.4 hours total. All dangerous scenes in all parameter spaces are found, and the test result is shown in fig. 7.
Claims (6)
1. A parallel acceleration test method for automatic driving simulation is characterized by comprising the following steps:
determining a parallel accelerated test architecture, dividing test areas and determining tasks of each level;
step two, allocating the execution quantity of the bottom layers of different small areas to ensure the test effect of the whole parameter space;
step three, determining the bottom performer test parameters in the small area to ensure the test effect in the small area;
step four, defining the bottom layer execution generation corresponding to the test result;
and step five, setting a test termination condition and interrupting the test process in time.
2. The parallel acceleration test method for the automatic driving simulation, according to claim 1, is characterized in that the specific method of the first step is as follows:
in the test process, a tested logic scene is regarded as a complete large area, and the large area of the logic scene parameter space is uniformly divided into a certain number of small areas; establishing a parallel accelerated test dangerous scene by designing a three-layer architecture; the three-layer architecture is respectively a top-layer manager, a secondary manager and a bottom-layer executor; the top manager is a scheduling unit at the top layer of the algorithm and is responsible for mutual scheduling among bottom executors which are responsible for different secondary managers; the secondary manager is a middle-layer scheduling unit in the algorithm and is responsible for planning the next working content of different bottom-layer executors in the small area, namely the parameters of the next test, and meanwhile, the secondary manager can mark the easily lost person in the responsible area range; the easily lost person is easy to leave a small area which is responsible for the current secondary manager due to low output, and is allocated to a bottom executive of other small areas by a top manager, and the output is the danger level of the generated scene; the bottom executor is a calculation unit for executing the test specifically, and receives the test parameter requirement sent by the secondary manager and performs the test.
3. The parallel acceleration test method for the automatic driving simulation as claimed in claim 2, wherein the specific method of the second step is as follows:
21) setting a top executive responsible for scheduling the bottom executive among different small areas, wherein the specific process comprises the following steps: monitoring the average output of the current iteration in each small area, calculating the average output of the whole large area in the same round, and commanding secondary managers in the area to allocate the easily lost people in the small area to go to the area with larger average output or the area generating extreme points for the small area with the average output lower than the average output of the whole area; the probability that a person who is easy to lose leaves different small production areas is as follows:
in the formula, piThe probability that the easily lost person in the ith cell leaves the cell is shown;average value of all bottom-layer executors output for j iteration; gijAverage yield in the ith cell for the jth iteration; j is the number of iterations; t is a regulating parameter which represents the strength of the top-level executor, the larger the value is, the larger the strength of the top-level manager is, the easier the secondary manager can listen to the deployment of the top-level manager, and the value is not less than 10; k is a pre-designed iteration number; n is the bottom layer executor number expected to be allocated to each small area by the top layer manager, and the value is not less than the initially allocated bottom layer executor number; n isijThe number of bottom-layer executors of the ith small area in the jth iteration is counted; n isijmaxIs the ithThe number of remaining active extreme values of the small region in the jth iteration; if a plurality of loss-prone persons exist in a region, the leaving probability of the loss-prone persons in one iteration is calculated sequentially, namely the leaving probability of the loss-prone person with the minimum yield in the current iteration is calculated firstly, whether the loss-prone person leaves is determined, and then the loss-prone person with the second minimum yield is calculated;
22) when all bottom executors in a cell run off, a top executor needs to define the cell as a development-value-free zone, seal the cell and forbid other bottom executors from entering the cell;
23) probability o of wandering off from the Small production region to a different region in round j +1 if there is anyiThe inflow calculation process is sequential calculation through the formula (2), and after each easily lost person flows in, the first term at the upper right side of the equation is changed;
aj=(aini-aend)·(K-j)/K+aend (3)
bj=(bini-bend)·(K-j)/K+bend (4)
cj=1-aj-bj (5)
in the formula, ninInputting the number of small areas; n isj *Is n of inputinThe total number of bottom performers of the small area in the jth round; gj *Is n of inputinTotal production of small regions in round j; lj *Is n of inputinThe total number of the maximum extreme values generated by the small region in the jth turn; a isj,bj,cjIn order to adjust parameters, the weights of the three sub-items are adjusted, the sum of the weights is 1, and the influence weight of the number of newly generated extreme points is larger and larger as iteration progresses; n isijFor the bottom layer of the ith small region in the jth iterationThe number of executives; gijAverage yield in the ith cell for the jth iteration; lijThe maximum extreme value number in the ith cell of the jth iteration is obtained;average value of all bottom-layer executors output for j iteration; a isiniIs an initial extreme value weight coefficient; a isendA weight coefficient for a termination extremum; k is a pre-designed iteration number; j is the current iteration number; biniIs an initial output weight coefficient; bendTo terminate the yield weight coefficient;
each new churner inflow changes the number of underlying executions of a region, requiring recalculation of the probability of flowing into a different region.
4. The parallel acceleration test method for the automatic driving simulation as claimed in claim 2, wherein the specific method of the third step is as follows:
31) the secondary manager is responsible for recording the maximum yield position, the minimum yield position, all tested parameter positions, all yields and corresponding turn numbers of each bottom-layer executor in the small area where the secondary manager is located, which are known up to now; each round, the secondary manager defines a plurality of convergence areas centered on the known maximum extremum in the area, and the bottom performers within the convergence areas, which are not at the maximum extremum, move at the following speed:
ωj=(ωini-ωend)·(K-j)/K+ωend (7)
vi0=rand(0,vmax) (8)
in the formula, xijThe test parameter position of the ith bottom-layer performer in the area in the jth iteration is determined; v. ofijThe speed of the ith bottom-layer performer in the jth iteration is obtained; dijIs a distance xijThe parameter location of the nearest maximum extremum; h isijStopping the position of the maximum output of the jth iteration for the ith bottom-layer executor; t is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86;taking 2.05 as a learning coefficient; omega is an inertia coefficient; omegainiTaking 0.9 as an initial inertia coefficient; omegaendTo terminate the inertia coefficient, 0.4 is taken; omegajThe velocity weight coefficient in the iteration process of the previous round is obtained; k is a preset iteration number; j is the current iteration number; v. ofi0A speed parameter for a first iteration; v. ofmaxIs a set maximum speed;
32) the convergence zone will increase with the increase of the test rounds, and the formula of the convergence zone range is:
rj+1=(2-T)·rj (10)
in the formula, all s meeting the conditions form a convergence region; u is different dimensions of the parameter space; k is all dimensions of the parameter space; r isuIs the axial length of the convergence region in the u-th dimension; r isjThe axial length of a convergence region of the j-th iteration; r is1Is the initial convergence region axial length; y is the maximum value of the parameters of different dimensions of the parameter space; e is the minimum value of the parameters of different dimensions of the parameter space; m is the number of all bottom layer executions; r isj+1The axial length of a convergence region of the j +1 th iteration; t is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86;
33) for the bottom performers at the non-maximum extreme outside the convergence region, the performers move at the following speeds:
in the formula, T is a compression coefficient, the speed of a bottom-layer executor of each iteration is controlled, and the value is 0.86; omegajThe velocity weight coefficient in the iteration process of the previous round is obtained; v. ofijThe speed of the ith bottom-layer performer in the jth iteration is obtained;taking 2.05 as a learning coefficient; h isijStopping the position of the maximum output of the jth iteration for the ith bottom-layer executor; x is the number ofijThe test parameter position of the ith bottom-layer performer in the area in the jth iteration is determined;
searching a nearest untested position near a current round working point or a parameter point obtained by calculation by a bottom-layer executor moving to a non-specified parameter position through the calculation for exploration; directly selecting the unexperienced parameter closest to the bottom performer at the maximum extreme value as the next experimental parameter;
34) the secondary manager marks the probability f that the i-th bottom performer in the small area becomes an vulnerable personiAs shown in equation (13); the first term of addition in the formula represents the comparison between the output of the bottom performer and the output of all the bottom performers, the second term of addition represents the comparison between the output of the bottom performer and the output of the bottom performers in the area, w is a crowding coefficient and is used for representing the distribution condition of a tested area in the area, if the Euclidean distance between the available nearest point and the calculated point is greater than 2 times of the Euclidean distance of the initial discrete step length, w is 1, otherwise, the value is 0;
wherein A, B is a regulation parameter, the sum of which is 1, and both of them are taken to be 0.5;average output of all bottom performers within 5 iterations; gi' is the average yield of the ith bottom performer over 5 iterations in the small area;average yield, w, for the bottom performers over 5 iterations in the regionijThe crowding coefficient of the ith bottom-layer performer in the small area in the j iteration is obtained; when the iteration times are less than 5, the average value of the operated iteration times is obtained; in addition thereto, forFor example, whether a specific bottom performer in the region is concerned about, that is, whether a loss-prone person enters or not, or the whole region condition in 5 iterations is concerned about, that is, the average yield of the latest 5 iterations is divided by 5;
since the vulnerable person does not necessarily leave the working area immediately, it is still possible to continue working in the area for a while if a vulnerable person finds a maximum point in the working phase of the area or exceeds the maximum point in the iteration processAndthe mark of the easy-to-lose person is lost and becomes the common bottom-layer executor again;
to prevent the vulnerable from frequent flowing between different areas, it is specified that an underlying actor will not be marked as vulnerable when the number of iterations in the current area is less than 5.
5. The parallel acceleration test method for the automatic driving simulation as claimed in claim 2, wherein the concrete method of the fourth step is as follows:
the bottom layer executor receives the parameter point position sent by the secondary manager, puts the parameter at the parameter point position into the simulation platform for simulation, records whether the test process is collided or not, if not, records the maximum longitudinal TTC in the whole test process-1I.e. time to collision, maximum transverse TTC-1The yield G of the bottom performer in this experiment was:
wherein collision is a collision signal, which represents the occurrence of collision; LTTC-1For maximum longitudinal TTC of test car in test process-1;HTTC-1For maximum lateral TTC of test vehicle in test process-1。
6. The parallel acceleration test method for the automatic driving simulation as claimed in claim 2, wherein the concrete method of the fifth step is as follows:
and (4) setting a test ending mark, reaching the set iteration number K, or finding no new extreme value in the continuous q rounds after a certain iteration number, wherein q is more than or equal to 2, and ending the test.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110478572.0A CN113514254B (en) | 2021-04-30 | 2021-04-30 | Parallel acceleration test method for automatic driving simulation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110478572.0A CN113514254B (en) | 2021-04-30 | 2021-04-30 | Parallel acceleration test method for automatic driving simulation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113514254A true CN113514254A (en) | 2021-10-19 |
CN113514254B CN113514254B (en) | 2022-05-17 |
Family
ID=78063669
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110478572.0A Active CN113514254B (en) | 2021-04-30 | 2021-04-30 | Parallel acceleration test method for automatic driving simulation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113514254B (en) |
Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20050196047A1 (en) * | 2004-02-03 | 2005-09-08 | Yuri Owechko | Object recognition system incorporating swarming domain classifiers |
CN107421752A (en) * | 2017-07-13 | 2017-12-01 | 同济大学 | A kind of intelligent automobile test scene accelerates reconstructing method |
CN108986503A (en) * | 2018-04-27 | 2018-12-11 | 北京智行者科技有限公司 | The dispatching method and system of automatic driving vehicle |
CN109992884A (en) * | 2019-04-01 | 2019-07-09 | 清华大学 | Automatic driving vehicle Function Appraising method based on worst scene search |
US20190339393A1 (en) * | 2018-05-03 | 2019-11-07 | GM Global Technology Operations LLC | Method and apparatus crosstalk and multipath noise reduction in a lidar system |
CN110795818A (en) * | 2019-09-12 | 2020-02-14 | 腾讯科技(深圳)有限公司 | Method and device for determining virtual test scene, electronic equipment and storage medium |
US10698802B1 (en) * | 2018-03-21 | 2020-06-30 | Cadence Design Systems, Inc. | Method and system for generating a validation test |
CN111539112A (en) * | 2020-04-27 | 2020-08-14 | 吉林大学 | Scene modeling method for automatically driving vehicle to quickly search traffic object |
CN111736951A (en) * | 2020-06-15 | 2020-10-02 | 深圳裹动智驾科技有限公司 | Simulation method for automatic driving, computer device, and storage medium |
CN111814308A (en) * | 2020-06-08 | 2020-10-23 | 同济大学 | Acceleration test system for automatic driving system |
CN111898211A (en) * | 2020-08-07 | 2020-11-06 | 吉林大学 | Intelligent vehicle speed decision method based on deep reinforcement learning and simulation method thereof |
CN112036001A (en) * | 2020-07-01 | 2020-12-04 | 长安大学 | Automatic driving test scene construction method, device and equipment and readable storage medium |
CN112528359A (en) * | 2020-08-06 | 2021-03-19 | 湖北民族大学 | City expansion simulation method based on bargaining model and ant colony optimization algorithm |
-
2021
- 2021-04-30 CN CN202110478572.0A patent/CN113514254B/en active Active
Patent Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20050196047A1 (en) * | 2004-02-03 | 2005-09-08 | Yuri Owechko | Object recognition system incorporating swarming domain classifiers |
CN107421752A (en) * | 2017-07-13 | 2017-12-01 | 同济大学 | A kind of intelligent automobile test scene accelerates reconstructing method |
US10698802B1 (en) * | 2018-03-21 | 2020-06-30 | Cadence Design Systems, Inc. | Method and system for generating a validation test |
CN108986503A (en) * | 2018-04-27 | 2018-12-11 | 北京智行者科技有限公司 | The dispatching method and system of automatic driving vehicle |
US20190339393A1 (en) * | 2018-05-03 | 2019-11-07 | GM Global Technology Operations LLC | Method and apparatus crosstalk and multipath noise reduction in a lidar system |
CN109992884A (en) * | 2019-04-01 | 2019-07-09 | 清华大学 | Automatic driving vehicle Function Appraising method based on worst scene search |
CN110795818A (en) * | 2019-09-12 | 2020-02-14 | 腾讯科技(深圳)有限公司 | Method and device for determining virtual test scene, electronic equipment and storage medium |
CN111539112A (en) * | 2020-04-27 | 2020-08-14 | 吉林大学 | Scene modeling method for automatically driving vehicle to quickly search traffic object |
CN111814308A (en) * | 2020-06-08 | 2020-10-23 | 同济大学 | Acceleration test system for automatic driving system |
CN111736951A (en) * | 2020-06-15 | 2020-10-02 | 深圳裹动智驾科技有限公司 | Simulation method for automatic driving, computer device, and storage medium |
CN112036001A (en) * | 2020-07-01 | 2020-12-04 | 长安大学 | Automatic driving test scene construction method, device and equipment and readable storage medium |
CN112528359A (en) * | 2020-08-06 | 2021-03-19 | 湖北民族大学 | City expansion simulation method based on bargaining model and ant colony optimization algorithm |
CN111898211A (en) * | 2020-08-07 | 2020-11-06 | 吉林大学 | Intelligent vehicle speed decision method based on deep reinforcement learning and simulation method thereof |
Non-Patent Citations (2)
Title |
---|
J. CHEN, P. ZHAO, H. LIANG AND T. MEI: "A Multiple Attribute-based Decision Making model for autonomous vehicle in urban environment", 《2014 IEEE INTELLIGENT VEHICLES SYMPOSIUM PROCEEDINGS》 * |
朱冰 等: "基于场景的自动驾驶汽车虚拟测试研究进展", 《中国公路学报》 * |
Also Published As
Publication number | Publication date |
---|---|
CN113514254B (en) | 2022-05-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111750887B (en) | Unmanned vehicle track planning method and system for reducing accident severity | |
EP2003080B1 (en) | Method and system for determining, for any instant in time, total peak power consumption for bank of elevator cars | |
US5612519A (en) | Method and apparatus for assigning calls entered at floors to cars of a group of elevators | |
JP4870863B2 (en) | Elevator group optimum management method and optimum management system | |
CN111125831B (en) | Track traffic train energy-saving speed track curve optimization method | |
KR100202720B1 (en) | Method of controlling multi elevator | |
CN115204466B (en) | International route planning method with band-pass restriction | |
CN110599772A (en) | Mixed traffic flow cooperative optimization control method based on double-layer planning | |
CN105608894B (en) | A kind of method and device for determining sudden congestion status | |
US11804140B2 (en) | Trajectory planning method and trajectory planning algorithm for an aerial vehicle | |
CN103824450B (en) | Based on the large-scale activity Special running layout of roads method of traffic behavior rule | |
CN112047212A (en) | Elevator operation control method, device, computer equipment and storage medium | |
CN113988408A (en) | Multi-objective planning-based artificial guidance method for passenger flow evacuation in subway station | |
CN112486200A (en) | Multi-unmanned aerial vehicle cooperative countermeasure online re-decision method | |
KR0178322B1 (en) | Elevator group control system | |
CN113514254B (en) | Parallel acceleration test method for automatic driving simulation | |
CN116797043A (en) | Auxiliary intelligent decision method for airspace running performance evaluation and sector division | |
CN106875681B (en) | A kind of traffic congestion tendency prediction technique and system | |
CN116734877A (en) | Robot dynamic obstacle avoidance method based on improved A-algorithm and dynamic window method | |
CN109345135B (en) | Method for determining key process in flow construction progress control | |
CN113326962B (en) | Subway train ATO speed curve prediction method based on BP neural network | |
Tran et al. | A model predictive control-based lane merging strategy for autonomous vehicles | |
CN109534117B (en) | Space three-dimensional interactive elevator calling device and working method | |
CN116341392A (en) | Multi-step ship lock ship dispatching control method | |
CN114442638B (en) | Desilting robot based on big data path planning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |