Planning with iFALCON: Towards A Neural-Network-Based BDI Agent Architecture

This paper presents iFALCON, a model of BDI (belief-desire-intention) agent that is fully realized as a self-organizing neural network architecture. Based on multichannel network model called fusion ART, iFALCON is developed to bridge the gap between a self-organizing neural network that autonomously adapts its knowledge and the BDI agent model that follows explicit descriptions. Novel techniques called gradient encoding are introduced for representing sequences and hierarchical structures to realize plans and the intention structure. This paper shows that a simplified plan representation can be encoded as weighted connections in the neural network through a process of supervised learning. A case study using the blocks world domain shows that an iFALCON agent can also do planning to solve problems when the knowledge is incomplete.