Policy Network
A neural network used in reinforcement learning and robot control to map observed states to actions. Conventional policy networks are difficult to compose directly for long-horizon tasks due to spatial misalignment.
Core Facts
Timeline (last 90 days)
BETS jointly optimizes the belief model (GDE), policy network, and value network during training, and the belief model's loss minimizes the discrepancy between the generated distribution and the true hidden state, commonly using negative log-likelihood or maximum mean discrepancy.
In traditional reinforcement learning and end-to-end paradigms, policy networks (π) generally cannot be directly concatenated; for example, separately trained grasping (π1), navigation (π2), and placing (π3) networks are extremely difficult to directly chain together to complete long-horizon tasks because their observation spaces, action spaces, and intermediate states are hard to align.
The output of a decision model is often a label or probability distribution over a finite action set, similar to a policy network in reinforcement learning, whereas generative models require autoregressive sampling over a large vocabulary.
All Facts (3)
The output of a decision model is often a label or probability distribution over a finite action set, similar to a policy network in reinforcement learning, whereas generative models require autoregressive sampling over a large vocabulary.
65%UnverifiedBETS jointly optimizes the belief model (GDE), policy network, and value network during training, and the belief model's loss minimizes the discrepancy between the generated distribution and the true hidden state, commonly using negative log-likelihood or maximum mean discrepancy.
50%UnverifiedIn traditional reinforcement learning and end-to-end paradigms, policy networks (π) generally cannot be directly concatenated; for example, separately trained grasping (π1), navigation (π2), and placing (π3) networks are extremely difficult to directly chain together to complete long-horizon tasks because their observation spaces, action spaces, and intermediate states are hard to align.
50%