Embedded Files
TL;DR:
TL;DR:
We evaluate the capacity of three multi-task Offline MARL sequence models, trained on large and diverse multi-task datasets, to generalise zero-shot to held-out test tasks. We find that increasing dataset diversity significantly improves generalisation, while simply increasing the dataset size for a fixed number of training tasks does not. We also uncover favourable generalisation gains with respect to scaling the sequence model size.
We evaluate the capacity of three multi-task Offline MARL sequence models, trained on large and diverse multi-task datasets, to generalise zero-shot to held-out test tasks. We find that increasing dataset diversity significantly improves generalisation, while simply increasing the dataset size for a fixed number of training tasks does not. We also uncover favourable generalisation gains with respect to scaling the sequence model size.
Environments:
Environments:
We evaluate across three challenging multi-agent environments, including LBF, RWARE and Connector.
We evaluate across three challenging multi-agent environments, including LBF, RWARE and Connector.
Connector
Connector
Level-Based Foraging (LBF)
Level-Based Foraging (LBF)
Robot Warehouse (RWARE)
Robot Warehouse (RWARE)
Visualising Learnt Mulit-Task Policy on RWARE:
Visualising Learnt Mulit-Task Policy on RWARE:
Page updated
Google Sites
Report abuse