MoWM: Mixture of World Models via Hypernetwork-Generated VSA Transitions

Key Points

• The goal is to create an architecture that uses multiple hypernetworks for generating transition functions based on visual-symbolic approaches.
• Developed a new architecture called MoWM with M hypernetworks.
• Implemented a learned gating mechanism to determine applicable world models based on situations.
• Focused on the use of VSA (Vector Symbolic Architectures) in the transition functions.
• Achieved improved adaptability of world models in various situations.
• Demonstrated effective selection between different transition functions based on environmental context.