Zhengyao Jiang
Zhengyao Jiang
Home
Publications
Light
Dark
Automatic
1
Efficient Planning in a Compact Latent Action Space
We propose a novel planning-based sequence modelling method that can scale to high dimensionality state-action space.
Zhengyao Jiang
,
Tianjun Zhang
,
Micheal Janner
,
Yueying Li
,
Tim Rocktäschel
,
Edward Grefenstette
,
Yuandong Tian
Optimal Transport for Offline Imitation Learning
We present an offline imitation learning based on optimal transport that demonstrates strong performance and sample efficiency
Yicheng Luo
,
Zhengyao Jiang
,
Samuel Cohen
,
Edward Grefenstette
,
Deisenroth Marc
Graph Backup: Data Efficient Backup Exploiting Markovian Data
We propose to treat the transition data of an MDP as a graph, and define a novel backup operator exploiting this graph structure. Comparing to multi-step backup, our graph backup method allows counterfactual credit assignment, and can reduce the variance that comes from stochastic environment dynamics.
Zhengyao Jiang
,
Tianjun Zhang
,
Robert Kirk
,
Tim Rocktäschel
,
Edward Grefenstette
Grid-to-Graph: Flexible Spatial Relational Inductive Biases for Reinforcement Learning
We proposed a principled and flexible framework to encode relational inductive bias using a relational graph. The relational inductive biases are crucial for the generalization of neural network models and are usually hard-coded in the neural architectures.
Zhengyao Jiang
,
Pasquale Minervini
,
Minqi Jiang
,
Tim Rocktäschel
Neural Logic Reinforcement Learning
To address interpretability and generalization of DRL, we propose a novel algorithm named Neural Logic Reinforcement Learning (NLRL) to represent the policies in reinforcement learning by first-order logic.
Zhengyao Jiang
,
Shan Luo
Cite
×