SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2405.17477 · ICML · 2024

OLLIE: Imitation Learning from Offline Pretraining to Online Finetuning

Sen Lin, Junshan Zhang, Sheng Yue, Xingyuan Hua, Ju Ren, Yaoxue Zhang

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 13 functions out of this paper's own repositories and ran 6 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
secury/optidice canonical 2 of 6
hansenhua/ollie-offline-to-online-imitation-learning alias 4 of 7
FunctionStatusWhere it lives
FullyConnectedQFunction Ran hansenhua/ollie-offline-to-online-imitation-learning/offline_to_online.py
pointer only (licence: NONE) · get_code("60820261a045e8c6")
ReparameterizedTanhGaussian Ran hansenhua/ollie-offline-to-online-imitation-learning/offline_to_online.py
pointer only (licence: NONE) · get_code("90b115e018d31278")
ReplayBuffer Ran hansenhua/ollie-offline-to-online-imitation-learning/offline_to_online.py
pointer only (licence: NONE) · get_code("edb756571764039b")
TanhGaussianPolicy Ran hansenhua/ollie-offline-to-online-imitation-learning/offline_to_online.py
pointer only (licence: NONE) · get_code("e4002bdf5440538b")
dice_dataset Ran secury/optidice/util.py
pointer only (licence: NONE) · get_code("d19706eb97928131")
generate_random_mdp Ran secury/optidice/mdp.py
pointer only (licence: NONE) · get_code("6047e3c9383e4164")
OTOF Not yet run hansenhua/ollie-offline-to-online-imitation-learning/offline_to_online.py
pointer only (licence: NONE) · get_code("0a4fc95528324a12")
Trainer Not yet run hansenhua/ollie-offline-to-online-imitation-learning/offline_to_online.py
pointer only (licence: NONE) · get_code("c3ae101599b97fa4")
boolean Not yet run secury/optidice/default_config.py
pointer only (licence: NONE) · get_code("3b81875ee24223a6")
init_module_weights Not yet run hansenhua/ollie-offline-to-online-imitation-learning/offline_to_online.py
pointer only (licence: NONE) · get_code("5c12e4fde6f862c0")
move Not yet run secury/optidice/example_fourrooms.py
pointer only (licence: NONE) · get_code("4a2bd7bd68b877ce")
policy_evaluation Not yet run secury/optidice/mdp.py
pointer only (licence: NONE) · get_code("f9ceec323249d2b9")
solve_MDP Not yet run secury/optidice/mdp.py
pointer only (licence: NONE) · get_code("a85083311a121189")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

In this paper, we study offline-to-online Imitation Learning (IL) that pretrains an imitation policy from static demonstration data, followed by fast finetuning with minimal environmental interaction. We find the naïve combination of existing offline IL and online IL methods tends to behave poorly in this context, because the initial discriminator (often used in online IL) operates randomly and discordantly against the policy initialization, leading to misguided policy optimization and unlearning of pretraining knowledge. To overcome this challenge, we propose a principled offline-toonline IL method, named OLLIE, that simultaneously learns a near-expert policy initialization along with an aligned discriminator initialization, which can be seamlessly integrated into online IL, achieving smooth and fast finetuning. Empirically, OLLIE consistently and significantly outperforms the baseline methods in 20 challenging tasks, from continuous control to vision-based domains, in terms of performance, demonstration efficiency, and convergence speed. This work may serve as a foundation for further exploration of pretraining and finetuning in the context of IL.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2405.17477")
get_code_for_paper("2405.17477")
have("2405.17477")

Connect an agent — have() is free.