Systems and methods for independent audit and assessment framework for AI systems
AssigneeArizona Board of Regents on behalf of Arizona State University
InventorsSiddharth Srivastava, Pulkit Verma
Life of this patent
Estimating the term…
Abstract
Example systems and methods for AI capability assessment include query-based assessment of sequential decision making agents (SDMAs) in stochastic settings with minimal assumptions on SDMA internals. In these examples, a new approach is presented for modeling the capabilities of black-box AI systems by using an active learning approach that can effectively interact with the black-box AI systems and learn an interpretable probabilistic model describing the capabilities of the black-box AI systems.
Claims8
A method for autonomous assessment of sequential decision-making systems in stochastic settings, comprising:
- accessing an input A, that is the software and/or integrated hardware constituting an agent of a sequential decision making (SDM) system capable of sequential decision making, or planning, and a simulator Sim that the agent A can interact with to execute its behavior;
- inducing a system that accepts users' specification of objectives and computes and executes the behavior necessary for achieving them,
- learning an output model defining capabilities of the SDM system using query-based autonomous capability estimation, including:
- defining M as a set of capability models expressible in terms of the set of interpretable concepts or predicates known to the user,
- discovering and learning true capabilities of the agent A as a model in M in a manner that clarifies to the user the agent A's true capabilities and the effects of executing them, by posing queries to the agent of the SDM system by:
- generating possible capability models for a specific predicate and location combination within M,
- utilizing a forward search to generate policy simulation queries for a randomly selected subset of the generated possible capability models, wherein the policy simulation queries output a policy to reach a state where the subset of generated possible capability models provide different responses from each other, and
- utilizing responses to the policy simulation queries as data to eliminate possible capability models inconsistent with A's true capabilities from the set of capability models M,
- wherein the output model as learned models the set of A's capabilities in stochastic settings.
Dependencies are inferred from the claim text.