We designed and implemented a prototype mechanism for agents to communicate with each other for the purpose of reward maximization. The agents can be interpreted as solving POMDP in an approximate manner, and are designed to use their actions, inferences, and communication in a purposive manner. We examined the validity of the design by writing a test program that runs on the prototype implementation. This mechanism is also a candidate for a computational model of human communication.
When responding to a person in a conversation, (1)Based on the "clause(WORD + conjunction)+will" of the story(sentence), convert the story into an image(result of viewing), (2)Using associative / hierarchical memory, you expand the content(image) of the talk and promote understanding, (3)Based on this, decide what you want to convey to the other person, (4)Using associative / hierarchical memory to extract the contents(images) to be conveyed to the other person, (5)Convert and sort into WORD, summarize as a sentence while forming a clause, convert them into audio information, and speak. The following technology is used to realize the above with an artificial brain. A)Multiplexed activation propagation in associative/hierarchical memory B)Hierarchical abstraction control(narrows the image to the upper-middle concept) C)Overall understanding and individual processing(Actions focusing on the object of attention while grasping the whole picture)
In this study, the idea and functionality of natural intelligence based on establishment of structural or functional adjunction between environmental systems of different systems of interest are presented based on category theoretical descriptions of adjunction and extension with an implementation example of ordered structure recognition in a multi-armed bandit problem based on a strategy inspired by Schubert calculus.