Communication and computation aware distributed machine learning for vehicular networks

Patent metadata

Field	Value
Publication number	US-12369186-B2
Application number	US-202217652145-A
Country	US
Kind code	B2
Filing date	Feb 23, 2022
Priority date	Feb 23, 2022
Publication date	Jul 22, 2025
Grant date	Jul 22, 2025

Abstract

A computer-implemented method is provided for training a global machine learning model using a learning server and a set of vehicle agents connected to roadside units (RSUs). It involves selecting agents, associating them with RSUs based on proximity, and transmitting training data and deadlines. Agents train models locally, which are then aggregated through RSUs to refine the global model to desired precision. The method includes steps of selecting vehicle agents from a pool of the vehicle agents connected to the RSUs, associating the selected vehicle agents and the RSUs respectively based on distances from the selected vehicle agents to the RSUs configured to provide measurements of the distances to the learning server, and transmitting a global model, a selected agent set and deadline thresholds in each global training round to the RSUs configured to transmit the global model and training deadlines to the selected vehicle agents.

First claim

We claim:

1. A computer-implemented method for training a global machine learning model using a learning server and a set of vehicle agents connected to roadside units (RSUs), wherein the method uses a processor coupled with a memory storing instructions implementing the method, wherein the instructions, when executed by the processor, carry out at steps of the method, comprising:
selecting vehicle agents from a pool of the vehicles connected to the RSUs, wherein the vehicle agents include on-board computer units and on-board sensors configured to collect local data through trajectories of the selected vehicle agents on roads;
associating the selected vehicle agents and the RSUs respectively based on distances from the selected vehicle agents to the RSUs configured to provide measurements of the distances to the learning server;
transmitting a global model w k , a selected agent set V k and deadline thresholds d cmp and deadline thresholds d thr in each global training round k to the RSUs configured to multicast the global model w k and training deadlines d v cmp to the selected vehicle agents, wherein the associated RSUs compute the training deadlines d v cmp of the corresponding selected vehicle agents, wherein the selected vehicle agents locally train the global model w k independently using the local datasets collected by the on-board sensors of the selected vehicle agents to generate locally trained models, wherein the deadline thresholds d cmp and d thr are determined in each global training round such that d cmp <d thr , wherein the deadline threshold d cmp is a deadline for the selected vehicle agents to finish locally training the global model and the deadline threshold d thr is a deadline for the selected vehicle agents to finish uploading the locally trained models, wherein with a given deadline threshold d thr , a total delay d v tot of a vehicle agent v is determined such that the total delay d v tot is less than the given deadline threshold d thr , i.e., d v tot ≤d thr ; and
aggregating the locally trained models from the selected vehicle agents via the associated RSUs to update the global model until the global training round reaches a pre-determined threshold.

2. The method of claim 1 , wherein with a given deadline threshold d cmp , a training deadline d v cmp for the vehicle agent v is determined using the given deadline threshold d cmp and global model distribution delay d v down such that a summation of the global model distribution delay and the training deadlines d v cmp is less than the given deadline threshold d cmp , i.e., d v down +d v cmp ≤d cmp .

3. The method of claim 1 , wherein the total delay d v tot of the vehicle agent v is a summation of the global model distribution d v down , the model training delay d v cmp , the uplink queuing delay d v q,up and the local model uploading delay d v up , i.e. d v tot =d v down +d v cmp +d v q,up +d v up .

4. The method of claim 3 , wherein the d v down is downlink global distribution delay of vehicle agent v calculated as

d
v
down

=

κ
×
min
⁢

{

T
:
κ
×

(

∑

t
_

=
1

T

⁢

C
v
dn

(

t
¯

)

)

≥
S

,

T
∈

ℤ
+

}

.

5. The method of claim 3 , wherein the d v up is uplink locally trained model uploading delay of vehicle agent v calculated as

d
v
up

=

κ
×
min
⁢

{

T
:
κ
×

(

∑

t
_

=
1

T

⁢

C
v
up

(

t
¯

)

)

≥
S

,

T
∈

ℤ
+

}

.

6. The method of claim 3 , wherein the d v q,up is uplink queueing delay of the locally trained model uploading delay of vehicle agent v calculated as

Assignees

• Mitsubishi Electric Res Laboratories Inc

Inventors

• Guo Jianlin
• Pervej Ferdous
• Kim Kyeong-Jin
• Parsons Kieran
• Orlik Philip
• Di Cairano Stefano
• Menner Marcel
• Berntorp Karl

Classifications

• G08G1/0141

  for traffic information dissemination · CPC title

• G08G1/0133

  for classifying traffic situation · CPC title

• G08G1/0116

  from roadside infrastructure, e.g. beacons · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06N20/00

  Machine learning · CPC title