Deep Transformer Q-Network
A Q-value estimation architecture for partially observable reinforcement learning that uses a Transformer encoder as a memory module. Self-attention models sequences of past observations, actions and rewards in order to infer hidden state.
Timeline (last 90 days)
A study from Northeastern University proposes Deep Transformer Q-Networks, using the Transformer architecture to address memory and credit assignment problems in partially observable reinforcement learning.
The study replaces the RNN as the memory module with a Transformer encoder, whose input is the sequence of the agent's observations, actions, and rewards over the past several steps, and whose output is used to estimate Q-values.
All Facts (2)
A study from Northeastern University proposes Deep Transformer Q-Networks, using the Transformer architecture to address memory and credit assignment problems in partially observable reinforcement learning.
50%UnverifiedThe study replaces the RNN as the memory module with a Transformer encoder, whose input is the sequence of the agent's observations, actions, and rewards over the past several steps, and whose output is used to estimate Q-values.
50%