IMAGINATION-BASED NEURAL AGENT NETWORKS
Patent Information
- Application Number
- DE602018084944
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2018-05-22
- Publication Date
- 2025-08-27
- Estimated Expiration
- 2038-05-22
AI Technical Summary
Reinforcement learning systems face challenges in efficiently interacting with real-world environments, particularly in continuous control tasks where there is no finite search tree to consider, and struggle to balance computational cost with task performance.
A neural network system comprising a controller, model, manager, and memory modules that selectively decides between executing actions and imagining consequences, using a model neural network to predict outcomes and a manager to balance planning and action, with training based on a loss function that includes computational cost.
The system improves interaction with real-world environments by reducing the need for extensive training data and computing resources while enhancing the ability to distinguish between similar states and handle unforeseen conditions, achieving better performance and resource efficiency.