A hybrid automated planning model based on Vector Embedding Representations and Transformer Architecture.
The Cognitive Generative Planner (CGP) is an innovative hybrid planning system that integrates symbolic and subsymbolic representations, combining the expressiveness of symbolic reasoning with the generalization power of neural networks.
By leveraging Ideas (symbolic structures) and their vectorized counterpart Vector Ideas, CGP enables creative, scalable, and adaptive planning. The system is driven by a Transformer-based neural architecture and enhanced by a symbolic search component: the Situated Beam Search (SBS) algorithm.
- Hybrid planning using Ideas + Vector Ideas
- Symbolic and subsymbolic knowledge integration
- Transformer-based neural network
- Novel Situated Beam Search Algorithm (SBS)
- Creative and explainable plan generation
- Evaluation in complex simulated environments
CGP was evaluated in dynamic simulation environments and demonstrated:
- High accuracy in valid plan generation
- High diversity in generated plans
- Lower computational cost than conventional planners
- Generalization across different planning contexts
- Plan validity accuracy
- Diversity metrics
- Beam size vs. performance
- Impact of temperature settings on generation behavior
- Python
- PyTorch
- Transformers
- Vector Embeddings
- Symbolic AI & Subsymbolic AI
- Knowledge Representation
- Cognitive System Toolkit (CST)
- Cognitive Generative Planner (CGP)
- Situated Beam Search (SBS)
- Transformer Architecture (Vaswani et al., 2017)
- Vector Embedding Representations
- Cognitive System Toolkit (CST)
- Deployment in real robotic environments
- Multi-agent planning extensions
- Integration with motivational and episodic memory systems
- Multimodal planning capabilities
This project is licensed under the MIT License.
Eduardo de Moraes Fróes
Ph.D. in Computer Engineering – UNICAMP
Advisor: Prof. Dr. Ricardo Ribeiro Gudwin