System and method for optimizing autonomous vehicle motion planner

By building models and using probability proxy functions or evolutionary algorithms to optimize the hyperparameters of the motion planner of autonomous vehicles, the problems of difficulty in selecting superparameters and insufficient generalization capabilities in the existing technology are solved, and safe and accurate decision-making in new scenarios are achieved , improves the vehicle's autonomous driving performance.

CN118922342APending Publication Date: 2024-11-08HUAWEI TECH CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202280093627.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2022-03-15
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Existing autonomous vehicle motion planners have difficulties in selecting and adjusting hyperparameters, especially the difficulty in determining the optimal set of hyperparameters that can generalize to new locations and scenarios, and often requires human intervention.

Method used

By constructing models to determine the improved hyperparameter set, using probability proxy functions or evolutionary algorithms to generate hyperparameter trial sets, and calculating new utility scores based on the true value result data, optimizing the hyperparameters of the motion planner to mimic human driving decisions.

Benefits of technology

The hyperparameter optimization of the autonomous driving vehicle motion planner is realized. The generated hyperparameter set can generate safe and accurate decisions in unknown scenarios, improve the vehicle's autonomous driving performance, and reduce the dependence on human intervention.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118922342A_ABST
    Figure CN118922342A_ABST
Patent Text Reader

Abstract

An apparatus (900) is described for determining an improved motion planner hyper-parameter (402). The apparatus (900) is configured to: receive a data pair (306), where the data pair (306) comprises a hyper-parameter set (402) and a utility score defining a utility of a motion planner (202) result (404) produced by the hyper-parameter set (402); providing a model, wherein the model defines a relationship between the hyper-parameter (402) and the utility score; generating a test hyper-parameter (402) using a guide target (304), wherein the guide target (304) is used to evaluate a quality of a hyper-parameter test set (402) from the model; determining a test result (404) of the motion planner based on the test hyper-parameter (402); a new utility score for the trial hyper-parameter (402) is determined (406) based on a comparison of the trial result (404) to truth data (408). Thus, the apparatus optimizes the currently deployed hyper-parameter (402) by comparing to the truth data to provide a hyper-parameter (402) that can produce an actual motion planned trajectory and is superior to the currently deployed hyper-parameter (402).
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to finding improved hyperparameters for autonomous vehicle motion planners, in particular using machine learning methods such as Bayesian optimization and / or evolutionary algorithms. Background Art

[0002] The goal of a reliable autonomous driving vehicle (ADV) is to make realistic and safe decisions for controlling the vehicle's motion, with or without real-time user input. These decisions are controlled by a motion planner in the ADV. However, an optimal motion planner is designed to output human-like decisions for all the motions of the car. Therefore, a well-designed motion planner should be designed not only to consider hard constraints such as avoiding collisions with road users, obstacles, and pedestrians, but also to make the output of the motion planner mimic the way humans drive. However, implementing a motion planner in this way is not easy.

[0003] The first challenge is to design a suitable architecture for the motion planner that can absorb the large amount of information obtained from the ADV's surrounding environment and process this information to generate appropriate motion planning decisions. The second challenge is how to train and / or run the motion planner after designing the motion planner architecture so that the motion planner can most effectively use the input information to generate accurate decisions.

[0004] The present invention is generally directed to the second difficulty. It is known in the art that motion planners use weights or hyperparameters that affect how the motion planner processes information from the surrounding environment. However, since the relationship between hyperparameters and motion output is very complex and generally unknowable, it is difficult to select or design an optimal set of hyperparameters for a motion planner. Therefore, for most given motion planners, it is generally not possible to determine the "optimal" set of hyperparameters analytically. For example, motion planners typically utilize hundreds or thousands of hyperparameters whose values ​​can be integers, continuous variables, categorical variables, Boolean values, etc. In addition, motion planner algorithms are typically complex and computationally expensive to run repeatedly during training. Therefore, even with a high-quality human-labeled decision set to determine the optimal hyperparameters, the task of selecting the optimal hyperparameters for deploying a motion planner remains challenging.

[0005] Furthermore, it is difficult to choose a set of hyperparameters that generalizes to new locations. In other words, the chosen hyperparameters should not "overfit", i.e., the chosen hyperparameters should produce accurate and safe decisions in a wide variety of locations and scenarios, not just for the types of locations / scenarios used in the data in which the motion planner was trained.

[0006] Due to these difficulties, motion planner hyperparameters usually need to be tuned by humans. However, several related methods for improving the performance of autonomous vehicles are known in the art.

[0007] In US2020 / 0150671, a system is disclosed that generates and scores trajectories based on a reward function using inverse reinforcement learning. The trajectory with the highest reward is then selected to control the ADV. However, before the "best" trajectory is selected, multiple different trajectories need to be determined. This invention does not directly disclose adjusting the parameters of the motion planner.

[0008] WO 2020 / 056331 A1 discloses a system for collecting training data, wherein the training data is collected from real ADVs and a neural network classifier is used to determine whether the training data is of sufficient quality. This invention does not disclose how to train hyperparameters, nor does it disclose the use of human-labeled decision data that is essentially considered to be correct.

[0009] A genetic algorithm is disclosed in CN 105946858 A that adjusts the parameters of an estimation model used to predict important attributes (e.g., longitudinal tire forces for downstream control methods). This invention does not disclose adjusting the parameters of a sampler that weights the trajectory to be selected based on the trajectory with the lowest cost.

[0010] A system is disclosed in CN108216250A, which is used to adjust parameters using a machine learning model to adaptively change ADV parameters based on instant feedback from passengers. Therefore, this invention involves real-time adjustment of parameters during and after deployment, and does not disclose the adjustment and selection of all hyperparameters before ADV deployment.

[0011] In "Hyperparameter Optimization in Black-box Image Processing using Differentiable Proxies", published in ACM SIGGRAPH 2019, optimization of the image processing component of the ADV system was disclosed. This therefore involves optimization of the processing of the input information, but optimization of a system for weighting different trajectories for vehicle control, where the image processing component remains unchanged, is not disclosed.

[0012] In "Hyperparameter Optimization using Grid Search for use in Monocular Depth Estimation in Cat Vehicles 2020 Technical Report", 2020, hyperparameter tuning of the depth estimation model is disclosed. Therefore, this involves tuning of a specific subset of hyperparameters, but does not disclose the general tuning of all hyperparameters of the motion planning system.

[0013] “Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles — NIPS 2017” discloses methods for quantifying prediction uncertainty in deep neural networks.

[0014] In “Hyperband: A NovelBandit Based Approach to Hyper parameter Optimization — JMLR 2018: Focus on a Different Problem”, we disclose multi-fidelity hyperparameter optimization of large machine learning models, where the training process can be temporarily stopped to modify parameters.

[0015] It would be advantageous to have an efficient scheme for selecting optimized hyperparameters for ADV motion planners, where the optimized hyperparameters outperform currently deployed human-selected suboptimal hyperparameters. It would also be advantageous to provide hyperparameters that generalize to unknown locations and scenarios. Summary of the invention

[0016] According to a first aspect, a device is provided for determining improved hyperparameters for use in an autonomous vehicle motion planner, the device comprising one or more processors and a memory, wherein the memory stores data in a non-transient form, the data defining a program code executable by the one or more processors to determine the improved hyperparameters. The device is used to: receive data comprising at least one data pair, wherein each data pair comprises a hyperparameter set and a utility score, wherein the utility score defines the utility of a motion planner result produced by the hyperparameter set; provide a model based on the at least one data pair, wherein the model defines a relationship between the hyperparameter set and a corresponding utility score; generate at least one hyperparameter test set using a guided target, wherein the guided target is used to evaluate the quality of the hyperparameter test set according to the model; determine a test result of the motion planner according to the hyperparameter test set and predefined trip data; determine a new utility score for the hyperparameter test set, wherein the new utility score is determined based on a comparison of the test result with the true value result data associated with the predefined trip data; generate a new data pair, wherein the new data pair comprises the hyperparameter test set and the new utility score.

[0017] Advantageously, by building a model based on existing hyperparameter data (e.g., hyperparameters used in currently deployed autonomous vehicles), and by calculating new utility scores for the experimental hyperparameters based on ground truth outcome data (e.g., human-labeled decisions), a new set of hyperparameters can be generated that outperforms the current hyperparameters and can also run a motion planner to produce actual results consistent with how humans drive. This is beneficial for safely integrating autonomous driving vehicles (ADVs) into real-world traffic environments. Furthermore, by generating experimental hyperparameters based on a model scored based on predefined trip data (preferably, the predefined trip data is real driving data obtained from real roads, and preferably also includes various environments), the improved hyperparameters can be advantageously applied to new, unknown scenarios that do not necessarily form part of the predefined trip data.

[0018] In some implementations, the model can be a probabilistic proxy function, and the guided objective is used to generate the at least one hyperparameter test set by fitting the probabilistic proxy function to one or more of the at least one data pair; and searching the hyperparameter input domain space based on sampling the probabilistic proxy function. Advantageously, probabilistic proxy models can efficiently guide the search for experimental hyperparameters because such proxy models typically run in much less time than static or dynamic motion planners. In addition, probabilistic proxy functions can guide the search for experimental hyperparameters by determining the prediction uncertainty of new data points (e.g., hyperparameter sets).

[0019] In some implementations, the probabilistic proxy function may be a Gaussian process model, while in other implementations, the probabilistic proxy function may be a Gaussian mixture model formed by combining multiple neural networks. Gaussian processes have the advantage of being very flexible and even applicable to complex functions such as motion planners. In addition, Gaussian processes can be easily initialized based on only one data pair and used to generate more data pairs in subsequent iterations. Gaussian mixture models (GMMs) formed by combining multiple neural networks provide an advantage in that even very noisy and / or complex functions can be modeled, i.e., this allows for effective exploration and exploitation of proxy functions.

[0020] In some implementations, the search of the hyperparameter input domain space is guided by an acquisition function, wherein the acquisition function is used to calculate the quality of the hyperparameter test set based at least in part on the predicted uncertainty of the value of the proxy function generated by the hyperparameter test set. The advantage of this is that the perceived value of the hyperparameter test set (i.e., the perceived value calculated by the model) can be offset by the predicted uncertainty of the value. Therefore, the acquisition function has the advantage of adjustability, for example, the trade-off between effective exploration and exploitation can be determined, so that the global maximum or global minimum of the proxy function can be determined.

[0021] In some implementations, the acquisition function includes one or more of the following functions: expected utility function, improved probability function, and upper confidence limit function. These functions can be easily used alone or in combination, and can be easily exchanged or interchanged throughout the process executed by the device.

[0022] In some implementations, the search of the hyperparameter input domain space is guided by an evolutionary algorithm. For example, the search of the hyperparameter test set using an acquisition function can itself be guided by an evolutionary algorithm. In other examples, a gradient-based approach can be used to guide the search for test hyperparameters using an acquisition function.

[0023] In some implementations, the model is the motion planner, and the guiding objective is used to generate the at least one hyperparameter test set using an evolutionary algorithm to randomly determine new data pairs by evaluating the utility scores of one or more new hyperparameter sets. In other words, the model can be the motion planner itself (preferably, a static motion planner), rather than a proxy function or approximation of the motion planner. Specifically, some examples can adopt direct sampling of the motion planner, for example, when the computational cost of the motion planner is not very high. The embodiment in which the model is a motion planner has an advantage, that is, since the motion does not approximate the relationship between the hyperparameter and the corresponding utility score, but directly calculates the relationship, therefore, the quality of the experimental hyperparameters determined by the model and / or guiding objective is accurate. Therefore, by directly querying the motion planner, the evolutionary algorithm can efficiently converge to the improved hyperparameters.

[0024] In some implementations, the test result is a vehicle trajectory, wherein the vehicle trajectory includes a plurality of vehicle motion decisions corresponding to at least one type of vehicle motion action. For example, the vehicle motion action may be a speed change such as acceleration or braking, a lateral planning decision such as steering or lane change, an indication, etc.

[0025] In some implementations, the utility score represents the accuracy of the trial results compared to the ground truth results data, wherein the ground truth results data includes human labeled vehicle motion decisions. Advantageously, using human labeled decisions enables the process to provide trial hyperparameters that, when used to run a motion planner, provide more realistic driving decisions (e.g., decisions that mimic human decisions). Thus, human labeled decisions have a benefit in that it is more likely to find hyperparameters that outperform currently deployed hyperparameters.

[0026] In some implementations, the utility score is calculated based on an objective function, wherein the objective function rewards correct decisions in the test results and / or penalizes incorrect decisions in the test results that were previously correctly determined based on an initial set of hyperparameter inputs in the received data. The inventors have discovered that it is beneficial to heavily penalize hyperparameter sets if they produce incorrect decisions that are correctly determined by the currently deployed hyperparameters. Providing this metric in the utility score calculation helps encourage the generation of improved hyperparameters and discourages the return of poorly performing test hyperparameters. In other words, a utility score determined in this manner has an advantage that the generated test hyperparameters should be at least as good as the hyperparameters in the received data.

[0027] In some implementations, the predefined trip data includes a plurality of trips, and the apparatus is further configured to: determine a plurality of test results using the hyperparameter test set, one test result for each trip; and determine the utility score based on the plurality of test results. For example, the predefined trip data may include data from real trips conducted in different cities and under different traffic conditions and driving environments. Using predefined trip data including a plurality of trips has an advantage that the generated hyperparameters can generalize well to unknown environments.

[0028] In some implementations, the apparatus is configured to determine the test results of the motion planner using a static simulator. A static simulator can still provide realistic test results, but operates more efficiently than a dynamic simulator.

[0029] In some implementations, the device is also used to iterate the steps of generating at least one hyperparameter test set and determining a new utility score, and the received data includes previously generated new data pairs. In this way, the device is used to build on the data so that subsequent iterations can produce more valuable hyperparameter test sets. Advantageously, because the received data contains more information as a basis for the model, the model can produce higher quality test hyperparameters after subsequent iterations. Further advantageously, it can be understood that the quality of the hyperparameter sets in the received data is irrelevant to the generation of high-quality hyperparameters, because each hyperparameter set has an associated score indicating the quality of each hyperparameter set. Therefore, the model essentially encodes in the relationship how to determine good test hyperparameters and how to avoid generating poor quality test hyperparameters.

[0030] According to another aspect of the present invention, a method is provided for determining improved hyperparameters for use in an autonomous vehicle motion planner. The method comprises: receiving data comprising at least one data pair, wherein each data pair comprises a hyperparameter set and a utility score, wherein the utility score defines the utility of the motion planner result produced by the hyperparameter set; providing a model based on the at least one data pair, wherein the model defines the relationship between the hyperparameter set and the corresponding utility score; generating at least one hyperparameter test set using a guided target, wherein the guided target is used to evaluate the quality of the hyperparameter test set according to the model; determining the test result of the motion planner based on the hyperparameter test set and predefined trip data; determining a new utility score for the hyperparameter test set, wherein the new utility score is determined based on a comparison of the test result with the true value result data of the predefined trip data; generating a new data pair, wherein the new data pair comprises the hyperparameter test set and the new utility score. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The present invention is described by way of example with reference to the accompanying drawings, in which:

[0032] Figure 1 A representation of environmental inputs used by an autonomous vehicle during a trip is shown;

[0033] Figure 2 A schematic diagram representing the architecture of a control system for an autonomous vehicle is shown;

[0034] Figure 3a and Figure 3b The iterative process performed to generate improved hyperparameter inputs is outlined separately, where the motion planner is treated as a black box;

[0035] Figure 4 Shows relative to Figure 3a and Figure 3b A more detailed iterative process that is performed to generate improved hyperparameter inputs, where the motion planner is treated as a black box;

[0036] Figure 5 Shows Figure 4 Part of the iterative process shown, where the solver uses Bayesian optimization;

[0037] Figure 6 Two approaches to performing the sampling process using Bayesian optimization are shown;

[0038] Figure 7 An exemplary method of the present invention is shown, which is used to directly sample a motion planner using an evolutionary algorithm, and some exemplary results of running the algorithm are also shown;

[0039] Figure 8 Shows Figure 4 An example of a portion of the iterative process shown, in which the quality of the motion planner output generated using an improved set of hyperparameter inputs is calculated;

[0040] Fig. 9 An example of an apparatus for performing the methods described herein is shown;

[0041] Fig.10 An exemplary method of the present invention for providing improved hyperparameters for use in an autonomous vehicle motion planner is shown;

[0042] Fig.11a and Fig.11b A graph showing the results of a dynamic autonomous driving simulation test conducted using improved hyperparameters generated by the method of the present invention. DETAILED DESCRIPTION

[0043] The present invention relates to an apparatus and method for determining an improved set of hyperparameters or weights for operating a motion planner of an autonomous driving vehicle (ADV).

[0044] Embodiments of the present invention are intended to solve the above problems by searching for hyperparameters that are better than the hyperparameters in the currently deployed ADV, for example, the hyperparameters of the currently deployed can be partially or completely manually selected / adjusted. In this context, "better than" should be understood to mean that the results of the motion planner (run by the improved hyperparameter set) contain safer and / or more accurate trajectories and / or contain more decisions that are considered to be correct. For example, the trajectory can be compared with a set of trajectory decisions that have been human-labeled so that they can be considered as "ground truth" data.

[0045] Embodiments of the method provided by the present invention use alone or in combination with algorithms such as Bayesian optimization or evolutionary algorithms to sample the huge domain space of hyperparameter inputs in order to find an improved hyperparameter set. In an example, the complexity of the motion planner itself makes it too expensive to repeat the operation (from the perspective of computing time). Therefore, an embodiment of the method uses a proxy function based on motion planner modeling to simulate the result of the motion planner. The proxy function is then sampled to obtain an improved hyperparameter set. In this way, the motion planner with high computational cost can only be run once (at most) in each iteration to test the accuracy / quality of the new hyperparameter set. In other examples, evolutionary algorithms can be used to directly sample the results of the motion planner, wherein the optimal hyperparameter set is used as the basis for generating a new hyperparameter set. The quality and / or utility of the new hyperparameter set can be objectively quantified using a suitable objective function (e.g., cost or reward function), which compares the output generated by running the motion planner using the trial hyperparameters with human labeled data (i.e., ground truth data) that are considered to be correct.

[0046] Thus, the present invention achieves results that provide improved hyperparameters that outperform the current best hyperparameters by sampling the results of the motion planner and comparing the hyperparameter trial set to the "ground truth" trajectory data, and does so in a time-efficient manner. Another benefit of obtaining hyperparameters that enable the motion planner to generate human-like decisions is that the motion of the ADV is more easily predicted by other humans, thereby allowing the ADV to better integrate into human-occupied roads.

[0047] Figure 1A computer rendering 100 of an autonomous vehicle on a journey is shown, showing various environmental conditions and objects. The data shown in the figure 100 represents a portion of the input used by the autonomous vehicle to guide the motion planner. The motion planner of the ADV uses a set of hyperparameters to indicate the relative importance of each environmental factor in the data acquired by the ADV in order to make motion decisions. The motion planner can contain hundreds or thousands of hyperparameters. Not all hyperparameters are independent, for example, there may be some degeneracy or overlap where two hyperparameters encode similar content. The motion planner uses different hyperparameters to affect different types of decisions, however, some hyperparameters are shared, i.e., affect multiple different decisions (or all decisions) made by the ADV.

[0048] Embodiments of the present invention are directed to simulating the decisions of a motion planner based on recorded real-world scenarios and improving the hyperparameters / weights used to run the motion planner so that the decisions output by the motion planner mimic expert (human) decisions. Preferably, the motion planner uses a static simulator as part of this process, however, in other embodiments, a more realistic dynamic simulator may be used.

[0049] Therefore, accurately selecting hyperparameters and preferably making the motion planner generate trajectories comparable to human driving styles is very beneficial to the safe operation of ADVs.

[0050] Figure 2 The architecture of a computing system 200 that controls an ADV is shown, and the location of a motion planner 202 implemented in the context of the system 200 is shown. The motion planner controls various aspects of the vehicle's motion, such as lateral planning involving obstacle avoidance and lane selection or cutting, and longitudinal planning involving vehicle acceleration and braking. There are generally four types of motion / trajectory controlled by the motion planner:

[0051] 1. General movement: that is, all interactions with the ADV environment.

[0052] 2. Lateral: Usually decisions involving steering, for example, decisions regarding when to change lanes (cut in), obstacle avoidance, lane discipline, etc.

[0053] 3. Longitudinal: Typically movements involving acceleration (i.e., using the accelerator) and deceleration (e.g., using the brakes), such as deciding when to give way to another vehicle at an intersection.

[0054] 4. Speed ​​planning: Plan the speed curve that ADV follows.

[0055] Throughout the present invention, it will be appreciated that the algorithm of the motion planner itself is predefined; the present invention is directed to methods of adjusting hyperparameters that affect how the motion planner operates. Therefore, it will be appreciated that the type of algorithm or architecture of the motion planner is irrelevant. Indeed, in embodiments of the present invention, the motion planner is viewed as a black box whose exact functionality is unknown, wherein the hyperparameters used to operate the motion planner are selected to optimize the output of the motion planner.

[0056] Figure 3a and Figure 3b An example is shown of performing the iterative process 300 to generate improved hyperparameter inputs, where the motion planner 202 is considered as a black box. Figure 3a The first step of a process 300 for generating improved hyperparameter inputs is shown, where the process starts with only one data point x1. Figure 3b Step i of the iterative process in a general case is shown, where the initial data set includes i data pairs.

[0057] exist Figure 3a In , the first data point x1 represents the initial hyperparameter set. Preferably, the first data point x1 comprises the hyperparameter set used in the currently deployed ADV, or comprises a hyperparameter set that has been verified (e.g., verified by a human expert) and is known to produce good results.

[0058] The initial input data x1 is used as input to the motion planner 202. The planner 202 is considered to be a black box whose functional / analytic form is unknown. Preferably, the motion planner used is a static simulator, run using real, predefined trip data. Preferably, multiple different trips are used to calculate the results of the input so that the improved set of hyperparameters generalizes more effectively to unknown positions. The motion planner 202 finally quantifies the quality of the hyperparameter input to generate a first data pair 306a, i.e.<x1,f(x1)> Preferably, the quality of the hyperparameter input is represented as a cost f(x1). This data pair forms part of the data 302 generated at each iteration.

[0059] At any given iteration, existing data 302, which includes all known data pairs 306a ... 306i, is provided to an agent 304. The agent 304 represents a model that searches the hyperparameter input domain space and generates a set of hyperparameter trials, i.e., hyperparameters that are predicted to provide improved results for the motion planner. The agent 304 uses a guidance goal to help it search for the set of trial hyperparameters, where the guidance goal is based on the existing data set (e.g., the first iteration).<x1,f(x1)> ). In the second iteration of the process, the agent 304 generates a hyperparameter trial set x2, which is fed into the motion planner 202. The agent then uses the second data pair 306b, i.e.<x2,f(x2)> , together with the first data pair, to provide other hyperparameter test sets.

[0060] In examples where the computational cost of motion planner 202 is high, it is advantageous to avoid running the motion planner. Therefore, preferably, agent 304 uses a simulation model of the motion planner (e.g., a proxy function when using Bayesian optimization) to generate the hyperparameter trial set. In other examples, the agent may use the motion planner algorithm itself to calculate the hyperparameter trial set x i quality (e.g., as part of an evolutionary algorithm).

[0061] It can be understood that in order to calculate the cost f(x) for a given set of hyperparameters i ), the motion planner first generates a motion planning output, i.e., a trajectory or decision set, based on which the cost can be calculated. Preferably, the cost function utilizes a set of human-labeled decisions based on the same predefined trip data, and the experimental hyperparameters are validated based on these human-labeled decisions. Therefore, the human-labeled decisions are considered to be ground truth data, or true result data.

[0062] Below is an example of a reward function used to validate the utility of a hyperparameter trial set against human-labeled decisions:

[0063]

[0064] in, The planner makes decision i for scenario j based on the adjusted parameters θ;

[0065] Decision i made by the planner for scenario j based on default parameters;

[0066] = the correct decision i made by humans for scenario j;

[0067] I[·] = Returns 1 if the internal statement is true, and 0 if it is false;

[0068] λ = importance weight.

[0069] Specifically, the value of λ is chosen to help influence the hyperparameter selection algorithm to retain previously correct decisions. The inventors have found that it is advantageous to provide a cost function that not only positively weights correct decisions, but also strongly penalizes using the default parameters. This would be a correct decision that was incorrectly determined using the hyperparameter test set. In other words, the search for good hyperparameters can be improved by strongly de-emphasizing hyperparameters that produce worse results than the currently deployed hyperparameters. Therefore, λ is selected to be a value greater than 1 to provide de-emphasis on new incorrect decisions, for example, where λ = 1.5 or 2.

[0070] Therefore, it will be apparent to those skilled in the art that the nature of the cost function and the functional form of the motion planner 202 have implications for the new data pair <x i ,f(x i )> has an impact on the generation of, and therefore affects the efficiency that can be used to determine improved hyperparameters. Therefore, in some embodiments, different forms of cost functions are used. Providing different forms of cost functions is within the purview of those skilled in the art. In some examples, the functional form of the cost function can be changed during process 300.

[0071] Figure 4 In more detail, the Figure 3a and Figure 3b The same iterative process 400 is performed in FIG. 4 , where the motion planner 202 is also regarded as a black box whose functional form is unknown.

[0072] Data pair set 306a...306i, namely<x1,f(x1)> ,<x2,f(x2)> … <x i ,f(x i )>, representing the starting data. The data is initially obtained from a data store 408, which may contain data related to the currently deployed ADV. Preferably, one of the data pairs contains a hyperparameter set x I , which represents the currently deployed hyperparameters. The data is passed to a solver 304 (also referred to as an "agent", such as agent 304 in FIG. 3 ), which generates trial search points 402 based on at least the existing data ( 306a . . . 306i ) and some form of guided goal, i.e., x i+1 The guide objective uses a known set of hyperparameters and their associated costs to guide the search for new data points 402. The trial data points 402 are then fed into the full motion planner 202. Preferably, the full motion planner 202 is a static trajectory simulator, i.e., a simulator that generates a set of trajectory decisions for each discrete point in a set of discrete points along a route of a predefined itinerary.

[0073] In some embodiments, the motion planner may also include a trajectory sampler that determines multiple potential trajectories that the ADV may take. For example, the sampler uses some knowledge of the road, static object interactions, vehicle dynamics, and modeled dynamics of other moving objects. The motion planner then uses the hyperparameters provided to it (i.e., 402, x i+1 ) to choose which trajectory should be taken to minimize the trajectory cost.

[0074] The motion planner 202 generates a decision set 404 associated with at least one predefined trip. In the example shown, <l1,...,l T > represents a horizontal decision set, <s1,...,s T > represents a set of velocity decisions. The results of these decisions are then fed into a cost function 406 which compares the trial motion planner output 404 with the human labeled data to calculate the cost f(x i+1 ).

[0075] As described above, the solver 304 can take a variety of different forms, such as Bayesian optimization and / or general algorithms, to generate new search points 402 that are predicted to produce improved motion planner decisions. Preferably, the motion planner 202 utilizes multiple predefined trips to create multiple decision sets 404, which are used to generate costs. In this way, the solver is more likely to produce improved trial hyperparameters that generalize to unknown driving positions, that is, the hyperparameters are not "overfit" to one specific type of trip or position.

[0076] Bayesian Optimization

[0077] Figure 5 An embodiment of a solver 304 of a process 400 using Bayesian optimization to generate new trial data points 402 is shown. The general approach of a Bayesian optimization routine is to statistically model some function that is considered to be a "black box" (i.e., the analytical form is unknown and / or the cost of direct sampling is too high). In this case, the motion planner represents the black box. The modeling approach can be divided into two major stages:

[0078] I. Learning Model 502. A function called a surrogate function is used as a substitute for the black-box motion planner 202. The goal of the surrogate model is to learn <x i ,f(x i)> so that new data points can be predicted. Therefore, a proxy function is sampled to replace the expensive black box function. There are many choices for the form of the proxy function, which is obvious to those skilled in the art. Figure 6 Two exemplary forms are described.

[0079] II. Maximizing the acquisition of function 504. Once the form of the surrogate function is established, the output range of the function can be explored to find the "optimal" input value. The acquisition function is used to explore the space of surrogate functions to find the input x that maximizes the output value of the surrogate function (or minimizes the cost, depending on what the surrogate function is modeling). i Therefore, the acquisition function can be considered as the guiding objective of the solver 304 using Bayesian optimization.

[0080] There are a large number of acquisition functions 504 that can be used alone or in combination to implement the acquisition function. Preferably, the acquisition function includes denoted as α EI Conceptually, when trying to minimize / maximize a function, the expected improvement provides a reward proportional to the improvement over the current maximum / minimum.

[0081] Multiple acquisition functions 504 may be used in combination. In the example shown, three functions are used: expected utility, improved probability, and upper confidence limit, the latter two denoted as α PI and α UCB .

[0082] The acquisition function above is defined as follows, assuming that f(x) is minimized, where f′ represents the lowest known value of f(x):

[0083] Improvement probability:

[0084] Expected improvement: u(x) = max(0,f′-f(x))

[0085] Thus, the acquisition function 504 described above is the expected utility as a function of x, for example:

[0086] in represents the current set of observations, also called priors, i.e. <x i ,f(x i )>.

[0087] The upper confidence limit is defined as: α UCB(x, β) = μ(x) - βσ(x), where μ(x) and σ(x) are the mean and variance, respectively, and β is a parameter for controlling the degree of exploration. Other acquisition functions are also known in the art and are suitable for use alone or in combination with the above functions. For example, other acquisition functions 504 include simple regret (SR), entropy search (ER), and knowledge gradient (KG).

[0088] Once the acquisition function determines the new search point 402, x i+1 , the new search point 402 is fed into the motion planner 202 and subsequently into the reward function 406 to quantify the quality of the new data point. In the case of Bayesian optimization, the new value of the proxy function f(x i+1 ) is called the posterior data point. Then, using the new posterior data <x i+1 ,f(x i+1 )>Update prior data.

[0089] Broadly speaking, the acquisition model is designed to consider both the predicted target value and the uncertainty surrounding the predicted target value when scoring a set of hyperparameters. The goal of the acquisition function 504 is to determine the minimum or maximum of the surrogate function. In order to efficiently search this area, it is advantageous to design the acquisition function to search beyond the local minimum or maximum range. This type of search, i.e., evaluating points with high prediction uncertainty to avoid being trapped in local minima / maxima, is called exploration. On the other hand, preferably, the acquisition function should be adjustable so as to give appropriate weight to the evaluation that returns the optimal value (i.e., the maximum or minimum). Searching in the area near the optimal value is called exploitation, for example, evaluating points with low mean.

[0090] Therefore, it is beneficial to design the acquisition to make an effective trade-off between exploration and exploitation. Various methods for implementing the acquisition function so that this trade-off can be adjusted will be apparent to those skilled in the art. For example, by UCB , μ(x) can be regarded as an exploitation term, and σ(x) can be regarded as an exploration term, which is controlled by the exploration parameter β. In addition, the expected improvement function inherently reflects the two considerations of exploration and exploitation.

[0091] In a preferred embodiment, the acquisition function 504 used in the present invention is constructed as:

[0092] max[α EI (·),α PI (·),α UCB (·)] T

[0093] It will be appreciated that the choice of whether to maximize or minimize the surrogate function depends on the form of the surrogate function: for example, whether the output of the surrogate function represents a reward (to be maximized) or a cost (to be minimized).

[0094] It will also be appreciated that once the metric to be optimized is defined using an acquisition function, there is a wide range of different methods for finding the minimum / maximum. In other words, once the acquisition function is determined, the search for new points is guided by the search function. For example, gradient-based methods known in the art can be applied to determine local maxima or minima in the surrogate function under the guidance of the acquisition function. Furthermore, the type of function used to guide the search is easily interchangeable.

[0095] Figure 6 Two alternative methods of constructing the surrogate function 502 for the Bayesian optimization solver 304 are shown.

[0096] In an alternative approach, the proxy function models the motion planner 202 using a Gaussian process 600. The use of Gaussian processes in Bayesian optimization schemes is known in the art. The true "black box" function f(x) is modeled as:

[0097]

[0098] This function takes the prior set As a condition,

[0099]

[0100] Where μ(x) represents the mean and k(x,x′) represents the variance.

[0101] In order to find the parameters of the Gaussian process that fits the current observation, an equation called the marginal likelihood is maximized. Since it is within the purview of those skilled in the art to implement this equation as part of a Gaussian process, the use and application of this equation will not be described in more detail. The marginal likelihood is defined as follows:

[0102]

[0103] A Gaussian process (GP) model is sampled using an acquisition function to provide predictions about new input values ​​402 that will maximize (or minimize, in the case where the surrogate function models a cost) the value of the GP. Any suitable acquisition function described above is suitable.

[0104] One advantage of using GP as a proxy function is that it allows quantifying the uncertainty of the prediction at any given point, for example, so that the likely value of an unknown search point can be evaluated. GP will initially produce inaccurate results. However, at each iteration, a new hyperparameter test set x is generated. i After that, the new hyperparameter trial set is then used to query the real model being simulated (motion planner 202), which gradually matures and produces more and more valuable outputs as Bayesian optimization proceeds (for example, a valuable set of hyperparameters causes the motion planner to output more accurate results than the currently deployed model).

[0105] Figure 6 The second alternative proxy function shown in FIG. 6 is a collection of neural networks (e.g., deep neural networks) 602. In this method, multiple neural networks are initialized independently and trained independently to perform well on the data. These neural networks are combined to create a Gaussian mixture model (GMM), where the predicted posteriors of the GMM are sampled using a collection function to determine new search points.

[0106] The inventors have found that in some cases, GP proxy models may produce complex and / or noisy functions that are difficult to optimize using acquisition functions. Therefore, GMMs formed from an ensemble of neural networks can be used for the most difficult cases, namely estimating the prediction uncertainty of new search points. GMMs using an ensemble of neural networks are generally more time-consuming than GP models, so when choosing between GP models and GMMs of neural networks, a trade-off between efficiency and accuracy is required. In principle, any number of neural networks can be combined; however, the inventors have found that more than 10 or about 15 neural networks can produce good results.

[0107] Furthermore, the types of surrogate functions can be interchanged during the Bayesian optimization process. For example, the process can be initialized with one surrogate function for an initial number of iterations and continue with a different surrogate function ( Figure 6 not shown).

[0108] It should be noted that an ensemble of neural networks may produce more accurate results than a GP proxy function, however, when using larger data sets A neural network is most effective when it is initialized. Therefore, in some embodiments, a more efficient GP model 600 may be initially used to construct a priori sets (e.g., where the data initially contains only one data pair).<x1,f(x1)> ), then GMM 602 is fitted to the data generated using GP and the iterative Bayesian optimization process continues. For example, once the data If the set of neural networks contains enough data pairs to initialize the set of neural networks, the GMM can continue the optimization process 400. For example, a Gaussian process surrogate can be used to obtain 1000 data points, and the GMM surrogate model (built using the set of neural networks) can be used in the remainder of the Bayesian optimization process to further refine the results. In general, when the data set becomes large, it may be advantageous to switch from a GP to a GMM surrogate function because the GMM is most effective when given a large data set, and the GMM can produce more accurate results than the GP model when given a large data set.

[0109] Alternatively, if the GP model starts to have difficulty producing improved hyperparameters (e.g., because the function is too noisy or complex), the GMM can take over.

[0110] When forming a GMM using an ensemble of neural networks, the following equation is used, where θ represents the set of hyperparameters:

[0111] (1)

[0112] (2)

[0113] (3)μ * (x) = M -1 ∑ m μθ m (x)

[0114] (4)

[0115] The above equation describes the following situation:

[0116] (1) The probability of y is equal to the average probability of y given each mixture in the GMM.

[0117] (2) The probability of y is equal to the normal distribution of the mean of each mixture and the combined variance of each mixture.

[0118] (3) Used to calculate the approximate center of GMM.

[0119] (4) Used to calculate the approximate variance of the GMM center.

[0120] Evolutionary Algorithms

[0121] As described above, the solver / agent 304 used as part of the optimization process can take many forms, and in some cases, the sample can directly query the motion planner. For example, this is appropriate when the computational cost of running the motion planner is not too high. In one embodiment, an evolutionary algorithm or a genetic algorithm can be used to directly sample the motion planner. However, the functional form of the motion planner does not need to be known for this purpose and can be regarded as a black box.

[0122] Alternatively, in examples where the computational cost of the motion planner is still too high to sample directly, the motion planner can be modeled using a proxy function in the manner of Bayesian optimization described above, and an evolutionary algorithm can be employed to optimize the acquisition function (i.e., rather than using a gradient-based approach). When an evolutionary algorithm is employed to optimize the acquisition function, the motion planner 202 can be used directly to calculate the "fitness".

[0123] Figure 7 is the general scheme of the evolutionary algorithm 700, where:

[0124] In step S100, control parameters of an evolutionary algorithm are initialized. In this example, the algorithm is a differential evolution (DE) algorithm.

[0125] S102 includes randomly initializing a population vector. For example, a random set of hyperparameters is selected to serve as a seed for the algorithm's starting point in the search for improved hyperparameters. Alternatively, when a "good" set of hyperparameters is already known (i.e., hyperparameters of the currently deployed known to perform well), this set can be used to generate an initial set of population vectors. Preferably, a large number of population vectors are typically initialized, for example, at least 50 or 100.

[0126] In S104, the fitness value of each vector is calculated, i.e., using a suitable objective function, such as a cost or reward function. For the avoidance of doubt, in this example, the calculation of the reward generally requires running the motion planner directly. In other examples, an approximation of the above cost function may be used, i.e., it is not necessary to run the motion planner 202, which is computationally expensive.

[0127] S106, S108 and S110 include mutation, crossover and selection steps, respectively. The goal after each iteration is usually to retain the best performing parameter set and remove the worst parameter set so that the mutation step generates new parameters only from the high-performance parameter set. Preferably, elitism is adopted so that the high-performance hyperparameter set is inserted into the next generation without undergoing mutation, which can provide favorable convergence of the genetic algorithm.

[0128] In S112, it is determined whether the termination condition has been met. If not, steps S106, S108 and S110 are repeated until the termination condition is met. For example, this condition may be met when the solution produced by the algorithm is better than the current best (i.e., currently deployed) hyperparameters, or when the performance of new hyperparameters no longer improves.

[0129] Once steps S106 , S108 , and S110 are terminated, in S114 , the vector with the smallest cost (conceptually equivalent to the largest fitness value) is selected as the solution.

[0130] Those skilled in the art will appreciate that there are many variations of genetic algorithms and that selecting an appropriate variation for use in the present invention is within the purview of those skilled in the art.

[0131] As described above, preferably, the objective function that scores (e.g., computes a cost or reward) for a given set of hyperparameters uses trip data from multiple different locations / trips to compute the score. Advantageously, this helps produce hyperparameters that generalize well regardless of road conditions, cities, road types, etc.

[0132] Figure 702 shows the results of running the DE evolutionary algorithm over multiple evaluation iterations. It can be seen that with an initial population of 20, the cost of optimizing the hyperparameter set can be reduced and optimized in, for example, 4000 evaluation steps. Figure 702 is the result of running the DE algorithm on a dynamic simulator (Kybersim) instead of a static simulator.

[0133] Figure 8 800 shows how the trip data from multiple different scenarios can be combined into a single score for a single hyperparameter set θ. The hyperparameters 402 (θ or x) used to run the motion planner i ) can be generated or obtained by any suitable method (including all methods of the present invention).

[0134] The multiple different scenarios represent different trips taken by the vehicle, preferably real trips providing realistic traffic and driving conditions. The data from each trip is processed separately by the motion planner 202, where the trajectory for each scenario is determined based on the same generated set of hyperparameters 402. In this scenario, preferably the motion planner is a static simulator. For each trip, the motion planner outputs a decision set 404 for each point in the static simulation. Although in Figure 6 Only lateral planning decisions are shown in , but any number of different vehicle motion decision types associated with different types of vehicle maneuvers or motions may be output.

[0135] The set of decisions from all scenarios 802 is then compared to the human labeled ground truth data 408, for example using the Cost(θ) function defined above, to score 406 the hyperparameters 402. Advantageously, for a given set of hyperparameters, the accuracy of the decisions associated with any number of different trips can be quantified and combined into a single target (cost) value.

[0136] In some embodiments, all available trip data may be used to calculate Cost(θ) in order to guide the construction of the agent model and / or the progress of the genetic algorithm. However, in other embodiments, it may be advantageous to run the hyperparameter improvement process 400 using only a subset of the trip data 802 and save the remaining data for cross-validation purposes. This is beneficial to verify that the process 400 does not overfit the data, i.e., is able to generate new hyperparameters that generalize well to new environments.

[0137] Fig. 9 An example of an apparatus 900 for implementing any method described herein is shown. The apparatus 900 can be implemented on an electronic device such as a laptop, a tablet computer, a smart phone, other mobile electronic devices, or a TV.

[0138] The apparatus 900 includes a processor 902 for processing a data set in the manner described herein. For example, the processor 902 may be implemented as a computer program running on a programmable device such as a central processing unit (CPU). The apparatus 900 includes a memory 904 for communicating with the processor 902. The memory 904 may be a non-volatile memory (e.g., permanent storage). The processor 902 may also include a cache ( Fig. 9 ), the cache may be used to temporarily store data from the memory 904. The apparatus 900 may include multiple processors 902 and multiple memories 904. The memory 904 may store data executable by the processor 902. The processor 902 may be used to operate according to a computer program stored in a non-transitory form on a machine-readable storage medium. The computer program may store instructions to cause the processor 902 to execute the method of the computer program in the manner described herein.

[0139] The processor may be implemented as a fixed logic circuit, for example as a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC) device. In addition, the apparatus may include multiple processors, the multiple processors being configured to run in parallel. It is within the purview of a person skilled in the art to implement at least a portion of the methods of the present invention so that the methods can be performed in parallel, for example, multiple results and utility scores for multiple predefined trips are easily calculated in parallel.

[0140] Specifically, the disclosed hyperparameter improvement device may include one or more processors (e.g., processor 902) and a memory 904, wherein the memory 904 stores data in a non-transient form, the data defining a program code executable by the one or more processors to implement a hyperparameter improvement device model, such as Fig.10 method steps.

[0141] Fig.10 An example of a method 900 for providing improved hyperparameters for use in an autonomous vehicle motion planner is outlined. S200 includes: receiving data including at least one data pair, wherein each data pair includes a hyperparameter set and a utility score, the utility score defining the utility of a motion planner result produced by the hyperparameter set. S202 includes: providing a model based on at least one data pair, wherein the model defines a relationship between a hyperparameter set and a corresponding utility score. S204 includes: generating at least one hyperparameter trial set using a guidance objective, wherein the guidance objective is used to evaluate the quality of the hyperparameter trial set according to the model. S206 includes: determining a trial result of the motion planner based on the hyperparameter trial set and predefined trip data. S208 includes: determining a new utility score for the hyperparameter trial set, wherein the new utility score is determined based on a comparison of the trial result with ground truth result data of the predefined trip data, and S210 includes: generating a new data pair including the hyperparameter trial set and the new utility score.

[0142] In a preferred example, the method is performed iteratively such that the data pairs generated in S210 are added to the received data used as input in S200.

[0143] As mentioned above, preferably, the motion planner 202 is a static simulator, as the operation of a dynamic driving simulator is generally computationally expensive and would therefore slow down the progress of the Bayesian optimization algorithm. However, the improved hyperparameters generated as part of the above method may be further validated, for example, in addition to cross-validation using the above static simulator.

[0144] The dynamic simulator can be used to validate the hyperparameters generated using the currently disclosed methods. This can be useful for assigning more accurate scores to hyperparameter sets and / or ensuring that the hyperparameter sets will produce safe results when deployed in the real world.

[0145] The results of such a dynamic simulation are shown in Figure 11. In this example, Kybersim is used to evaluate the hyperparameters by simulating different driving positions. Using a dynamic simulator such as Kybersim provides the benefit that more realistic (e.g., real-world) performance of the selected ADV parameters can be estimated.

[0146] Kybersim is a well-known industrial simulator for autonomous vehicles. The selected "best" set of hyperparameters output by one of the above embodiments is used as input. Given a set of parameters, a dynamic simulator is run so that objects can interact with vehicles (e.g. other (moving) vehicles) or pedestrians. Thus, one can better understand how the parameters selected by the agent (solver 304) will respond when deployed in the real world.

[0147] 1100 and 1102 respectively show the error and error change of the ongoing Kybersim simulation after multiple iterations. 1104 shows the improvement of the travel time of the Kybersim simulated trip after multiple iterations.

[0148] By employing the results of the Kybersim simulator shown and the hyperparameter optimization method described herein, the inventors were able to achieve a relative speedup of 60% in various scenarios.

[0149] Applicants hereby disclose separately each individual feature described herein and any combination of two or more such features. With the common knowledge of those skilled in the art, such features or combinations can be implemented as a whole based on this specification, without considering whether such features or combinations of features can solve any problems disclosed herein, and without limiting the scope of the claims. Applicants indicate that various aspects of the present invention can be composed of any such individual features or combinations of features. In view of the above description, it will be obvious to those skilled in the art that various modifications can be made within the scope of the present invention.

Claims

1. A device (300, 400, 900), It is characterized in that The apparatus (300, 400, 900) is used to determine improved hyperparameters (402) for use in an autonomous vehicle motion planner (202), the apparatus comprising one or more processors (902) and a memory (904), wherein the memory (904) stores data in a non-transitory form, the data defining program code executable by the one or more processors to determine the improved hyperparameters (402), the apparatus being configured to: Receiving data (302) comprising at least one data pair (306), wherein each data pair comprises a hyperparameter set and a utility score, the utility score defining the utility of a motion planner result produced by the hyperparameter set; providing a model (304) based on the at least one data pair, wherein the model defines a relationship between the set of hyperparameters (306) and corresponding utility scores; generating at least one hyperparameter test set (402) using a guided objective, wherein the guided objective is used to evaluate the quality of the hyperparameter test set according to the model (304); Determining a test result (404) of the motion planner (202) based on the hyperparameter test set (402) and the predefined travel data (802); determining (406) a new utility score for the set of hyperparameter trials, wherein the new utility score is determined based on a comparison of the trial results with ground truth result data (408) associated with the predefined trip data (802); A new data pair (306b, 306i) is generated, wherein the new data pair (306b, 306i) includes the hyperparameter trial set and the new utility score.

2. The device according to claim 1, It is characterized in that The model is a probabilistic proxy function (304), and the guided objective is used to generate the at least one hyperparameter trial set in the following manner: fitting the probabilistic proxy function to one or more of the at least one data pair; Based on the sampling of the probability proxy function, a hyperparameter input domain space is searched.

3. The device according to claim 2, It is characterized in that The probabilistic proxy function is a Gaussian process model (600).

4. The device according to claim 2, It is characterized in that The probabilistic proxy function is a Gaussian mixture model (602) formed by combining multiple neural networks.

5. The device according to any one of claims 2 to 4, It is characterized in that The search of the hyperparameter input domain space is guided by an acquisition function (504), wherein the acquisition function (504) is used to calculate the quality of a hyperparameter test set (402) based at least in part on a predicted uncertainty of a value of the surrogate function produced by the hyperparameter test set.

6. The device according to claim 5, It is characterized in that The acquisition function (504) includes one or more of the following functions: an expected utility function, an improved probability function, and an upper confidence limit function.

7. The device according to any one of claims 2 to 4, It is characterized in that The search of the hyperparameter input domain space is guided by an evolutionary algorithm (700).

8. The device according to claim 1, It is characterized in that The model (304) is the motion planner (202), and the guiding objective is to generate the at least one hyperparameter trial set (402) using an evolutionary algorithm (700) to randomly determine new data pairs (306i) by evaluating the utility scores of one or more new hyperparameter sets.

9. The device according to any one of the preceding claims, It is characterized in that The test result is a vehicle trajectory (404), wherein the vehicle trajectory (404) includes a plurality of vehicle motion decisions corresponding to at least one type of vehicle motion maneuver.

10. The device according to any one of the preceding claims, It is characterized in that The utility score represents the accuracy of the test results compared to the ground truth results data (408), wherein the ground truth results data includes human-labeled vehicle motion decisions.

11. The device according to claim 10, It is characterized in that The utility score is calculated based on an objective function, wherein the objective function rewards correct decisions in the trial results and / or penalizes decisions in the trial results that are incorrect and were previously correctly determined based on an initial set of hyperparameter inputs in the received data.

12. The device according to any one of the preceding claims, It is characterized in that The predefined trip data includes a plurality of trips (802), and the apparatus is further configured to: Determine a plurality of test results using the hyperparameter test set, one test result for each trip; The utility score is determined based on the plurality of test results.

13. The device according to any one of the preceding claims, It is characterized in that The apparatus is used to determine the test results of the motion planner using a static simulator (202).

14. The device according to any one of the preceding claims, It is characterized in that The apparatus is further configured to iterate the steps of generating at least one hyperparameter trial set and determining a new utility score, wherein the received data includes previously generated new data pairs.

15. A method, It is characterized in that The method is for determining improved hyperparameters for use in an autonomous vehicle motion planner (202), the method comprising: Receiving data (302) comprising at least one data pair (306), wherein each data pair comprises a hyperparameter set and a utility score, the utility score defining the utility of a motion planner result produced by the hyperparameter set; providing a model based on the at least one data pair (304), wherein the model defines a relationship between the set of hyperparameters and corresponding utility scores; generating at least one hyperparameter test set using a guided objective (402), wherein the guided objective is used to evaluate the quality of the hyperparameter test set according to the model; Determining a test result (404) of the motion planner based on the hyperparameter test set and the predefined travel data (802); determining (406) a new utility score for the hyperparameter test set, wherein the new utility score is determined based on a comparison of the test results with ground truth result data (408) of the predefined trip data (802); A new data pair is generated, wherein the new data pair includes the hyperparameter test set and the new utility score.

Citation Information

Cited By

  • Hyper-parameter tuning method and device for automatic driving closed-loop simulation, equipment and medium

    CN120871662A