US 12,393,581 B2
Method and system for specifying an active inference based agent using natural language
Capm Petersen, Los Angeles, CA (US); Conor Heins, Berlin (DE); Mahault Albarracin, Montreal (CA); Riddhi Pitlya, Oxford (GB); Tim Verbelen, Dendermonde (BE); Alexander Tschantz, West End (AU); Axel Constant, Montreal (CA); Christopher L. Buckley, Brighton (GB); Alexander Bernard Kiefer, Arverne, NY (US); Karl Friston, Rickmansworth (GB); Steven Swanson, Culver City, CA (US); Gabriel Rene, Los Angeles, CA (US); and Tommaso Salvatori, Vienna (AT)
Filed by Verses AI, Inc., Vancouver (CA)
Filed on Jul. 12, 2024, as Appl. No. 18/770,654.
Claims priority of provisional application 63/513,322, filed on Jul. 12, 2023.
Prior Publication US 2025/0021548 A1, Jan. 16, 2025
Int. Cl. G06F 16/2452 (2019.01); G06F 16/22 (2019.01); G06F 16/25 (2019.01)
CPC G06F 16/24522 (2019.01) [G06F 16/2237 (2019.01); G06F 16/258 (2019.01)] 10 Claims
OG exemplary drawing
 
1. A system for specifying artificial intelligence agent behavior using natural language user queries and inferring action policies of artificial intelligence agents for use in a physical or virtual device, comprising:
a processor on a computer system,
at least one storage medium in communication with the processor, which at least one storage medium includes code defining:
an agent model,
Create, Read, Update, and Delete (CRUD) functionality configured to communicate in natural language with
software that translates natural language into vectorized entities encoded in a knowledge graph implementing the Hyperspatial Modeling Language (HSML),
a hyperspatial Modeling Language (HSML) that defines a special implementation of a knowledge graph, referred to herein as an HSML graph, whose syntax defines entities in physical or virtual space that the agent interacts with, each entity comprising an entity type, and 4 entity subtypes that include (i) schema (ii) vector space, (iii) links, (iv) datalinks, wherein the schema contains a vectorization field for receiving the vector of the vectorized entity that is represented by the HSML knowledge graph, and wherein the vector space describes the structure of the vectors for the vectorized entity, and wherein the link defines the relationship between two or more entities, the HSML thus defining the syntax for translating entities described in natural language into vectorized representations of those entities expected by the HSML graph,
a translation software that uses the syntax defined by the HSML to translate the natural language description of entities into vectorized representations, wherein the translation software is a Large Language Model (LLM), or any software capable of transforming natural language into vectorized representations in accordance with the syntax defined by the modelling language of HSML, and
active inference agent software capable of computing a numerical value called
expected free energy based on a model whose parameters map the vectorized entities stored in the HSML graph, and that scores the value of action policies available to the agent,
the system further including the device that is controlled by the agent model.