Autonomous vehicle model training and validation using low-discrepancy sequences
Autonomous vehicle model training and validation using low-discrepancy sequences may include: generating a low-discrepancy sequence in a multidimensional space comprising a plurality of multidimensional points; mapping each sample of a plurality of samples of a data corpus to a corresponding entry in the low-discrepancy sequence, wherein each sample of the plurality of samples comprises one or more environmental descriptors for an environment relative to a vehicle and one or more state descriptors describing a state of the vehicle; selecting, from the data corpus, a training data set by selecting, for each multidimensional point of the low-discrepancy sequence having one or more mapped samples, a mapped sample for inclusion in the training data set; and training one or more models used to generate autonomous driving decisions of an autonomous vehicle based on the selected training data set.
1. A method comprising:
generating a low-discrepancy sequence in a multidimensional space comprising a plurality of multidimensional points;
mapping each sample of a plurality of samples of a data corpus to a corresponding entry in the low-discrepancy sequence, wherein each sample of the plurality of samples is encoded as a record comprising a plurality of field-value pairs, wherein the plurality of field-value pairs comprises one or more environmental descriptors for an environment relative to a corresponding vehicle and one or more state descriptors describing a state of the corresponding vehicle, wherein each sample is generated by and received from the corresponding vehicle;
selecting, from the data corpus, a training data set by selecting, for each multidimensional point of the low-discrepancy sequence having one or more mapped samples, a mapped sample for inclusion in the training data set; and
training one or more models used to generate autonomous driving decisions of an autonomous vehicle based on the selected training data set.
2. The method of claim 1 , wherein each sample of the plurality of samples comprises a plurality of continuous fields each corresponding to a dimension of the multidimensional space.
3. The method of claim 2 , wherein each sample in the data corpus further comprises one or more categorical fields.
4. The method of claim 3 , wherein selecting the training data set comprises selecting, for each combination of categorical field values of the one or more categorical fields, a subset of the training data set from a subset of the data corpus corresponding to a particular combination of categorical field values.
5. The method of claim 1 , wherein mapping each sample of the plurality of samples the corresponding entry in the low-discrepancy sequence comprises identifying, for each sample of the plurality of samples, as the corresponding entry, a nearest entry in the low-discrepancy sequence.
6. The method of claim 1 , further comprising receiving a request for a number of samples from the data corpus, wherein generating the low-discrepancy sequence comprises generating the low-discrepancy sequence as having a number of multidimensional points based on the number of samples indicated in the request.
7. The method of claim 1 , further comprising:
identifying at least one entry in the low-discrepancy sequence without a mapped sample; and
generating, based on the at least one entry, at least one synthetic training data entry for inclusion in the training data set.
8. The method of claim 7 , wherein identifying the at least one entry comprises:
identifying a sparsely mapped portion of the low-discrepancy sequence; and
identifying the at least one entry in the sparsely mapped portion of the low-discrepancy sequence.
9. The method of claim 8 , wherein identifying the sparsely mapped portion of the low-discrepancy sequence is based on one or more density thresholds.
10. The method of claim 9 , wherein the one or more density thresholds are based on one or more of: a particular type of model of the one or more models or a particular dimension in the multidimensional space.
11. The method of claim 1 , wherein the low-discrepancy sequence comprises a Sobol sequence.
12. An apparatus comprising at least one processor and memory storing instructions that, when executed, cause the at least one processor to perform steps comprising:
generating a low-discrepancy sequence in a multidimensional space comprising a plurality of multidimensional points;
mapping each sample of a plurality of samples of a data corpus to a corresponding entry in the low-discrepancy sequence, wherein each sample of the plurality of samples is encoded as a record comprising a plurality of field-value pairs, wherein the plurality of field-value pairs comprises one or more environmental descriptors for an environment relative to a corresponding vehicle and one or more state descriptors describing a state of the corresponding vehicle, wherein each sample is generated by and received from the corresponding vehicle;
selecting, from the data corpus, a training data set by selecting, for each multidimensional point of the low-discrepancy sequence having one or more mapped samples, a mapped sample for inclusion in the training data set; and
training one or more models used to generate autonomous driving decisions of an autonomous vehicle based on the selected training data set.
13. The apparatus of claim 12 , wherein each sample of the plurality of samples comprises a plurality of continuous fields each corresponding to a dimension of the multidimensional space.
14. The apparatus of claim 13 , wherein each sample in the data corpus further comprises one or more categorical fields.
15. The apparatus of claim 14 , wherein selecting the training data set comprises selecting, for each combination of categorical field values of the one or more categorical fields, a subset of the training data set from a subset of the data corpus corresponding to a particular combination of categorical field values.
16. The apparatus of claim 12 , wherein mapping each sample of the plurality of samples the corresponding entry in the low-discrepancy sequence comprises identifying, for each sample of the plurality of samples, as the corresponding entry, a nearest entry in the low-discrepancy sequence.
17. The apparatus of claim 12 , wherein the steps further comprise receiving a request for a number of samples from the data corpus, wherein generating the low-discrepancy sequence comprises generating the low-discrepancy sequence as having a number of multidimensional points based on the number of samples indicated in the request.
18. The apparatus of claim 12 , wherein the steps further comprise:
identifying at least one entry in the low-discrepancy sequence without a mapped sample; and
generating, based on the at least one entry, at least one synthetic training data entry for inclusion in the training data set.
19. The apparatus of claim 18 , wherein identifying the at least one entry comprises:
identifying a sparsely mapped portion of the low-discrepancy sequence; and
identifying the at least one entry in the sparsely mapped portion of the low-discrepancy sequence.
20. The apparatus of claim 19 , wherein identifying the sparsely mapped portion of the low-discrepancy sequence is based on one or more density thresholds.
21. The apparatus of claim 20 , wherein the one or more density thresholds are based on one or more of: a particular type of model of the one or more models or a particular dimension in the multidimensional space.
22. The apparatus of claim 12 , wherein the low-discrepancy sequence comprises a Sobol sequence.