11 results about "Formal methods" patented technology

The invention provides a mathematical proposition automatic formalization method and device based on reinforcement learning, and relates to the technical field of mathematics, and the method comprises the steps: obtaining a to-be-formalized mathematical proposition expressed by a natural language; inputting a mathematical proposition to be formalized into the target automatic formalized model to generate formalized language expression; according to the target automatic formalized model, after a sample data set of a mathematical proposition is input into an initial automatic formalized model to obtain formalized language expression corresponding to the sample data set, a reward signal is generated based on the formalized language expression corresponding to the sample data set through a grammar checker and a semantic consistency check model; and a group relative strategy optimization algorithm is adopted to carry out multi-round iteration updating on the model parameters of the initial automatic formalization model by using the reward signal. According to the method, the requirement for training data is greatly reduced, the training efficiency is improved, the training effect of an automatic form model is improved, and then the accuracy of automatic formalization of the mathematical proposition is improved.

The invention relates to the technical field of software engineering and artificial intelligence, and discloses a high-credibility programming agent system and method based on a large model and a formalization method. A high-credibility programming agent system based on a large model and a formalization method comprises an agent core module used for providing common perception, memory, decision and execution capabilities for various professional agents; the multi-agent cooperation module is constructed on the agent core module and comprises a coding agent, a form analysis agent and a form verification agent; the coding agent is responsible for generating, complementing, reconstructing and optimizing program codes based on natural language requirements or contexts. The system realizes full-process automation from demand understanding to code generation, analysis, verification and restoration through autonomous planning and multi-task execution capability of the intelligent agent, reduces manual intervention links, and remarkably shortens the development period.

This invention discloses a live-cell analysis method and system based on a large model and formal verification. The method includes context injection and intent reasoning, receiving natural language and analyzing it using a large language model to infer the required algorithm modules and output candidate configuration files. Formal logical verification involves reading the candidate configuration files and converting them into standardized logical constraint code, which is then verified by a solver. Conflict self-healing and iteration are implemented: if the solver verification fails, the conflict core is identified, and the connection method of the candidate configuration file or algorithm module is modified according to an arbitration strategy, while updating the logical constraint code until the solver verification passes. An analysis pipeline is instantiated and constructed: if the solver verification passes, an executable analysis pipeline is built based on the final determined candidate configuration files for live-cell analysis. Analysis code is generated using a large language model and logically verified using formal methods to analyze live-cell microscopic images.

The application discloses a kind of based on formalization method's trusted machine learning model explanation method and device, it is related to artificial intelligence technical field, the method includes obtaining target machine learning model and its input characteristic value;Input characteristic value is compiled using formalization method, to obtain continuous variable;Continuous variable is discretized, to obtain the input characteristic value after discretization;The logic variable compilation of characteristic value after discretization is carried out, to obtain the variable after formalization compilation;According to the decision rule of model, the logic expression after formalization compilation is simplified, to obtain the rule reason of the expression of implication, that is, decision;Rule reason is obtained using natural language representation, to obtain the explanation result of model.The application uses formalization method to explain trusted machine learning model, compared with traditional explanation method, can more accurately reveal the internal logic of model decision, to improve the reliability of model explanation.

The embodiment of the present disclosure discloses a method and device for formal analysis of autonomous positioning of a UAV. The method comprises: obtaining the same name image point coordinates, a preset homogeneous coordinate conversion matrix, a continuous homogeneous coordinate conversion matrix and a coordinate conversion formula, establishing a high-order logic expression of a ground target point in a geodetic coordinate system based on a high-order logic language; constructing a corresponding formal proposition based on the high-order logic expression, the formal proposition describing attribute constraint conditions in the high-order logic expression; verifying the formal proposition based on a theorem prover to determine whether the high-order logic expression meets the attribute constraint conditions. The high-order logic expression of the UAV for determining the non-homogeneous coordinates of the ground target point in the geodetic coordinate system is analyzed by the formal method, and the attribute constraint conditions that must be met by the high-order logic expression are verified, thereby ensuring the accuracy of the analysis result.

Methods for more secure mobile network communications are disclosed. Specifically, details involving natural language processing (NLP) based auto formal modeling of protocols and specifications with large language models (NLP) are provided. Methods for formal and fuzzing amplification for fuzz testing to detect vulnerabilities are also disclosed. Furthermore, solutions are provided to identified vulnerabilities in existing 5G infrastructures. Also disclosed is a digital twin fuzzing framework.

The invention discloses a mathematical automatic formalization method based on a large language model, which comprises the following steps of: initializing a system, setting an iteration counter t to be 1, and setting a historical record H as an empty set; an automatic formalization generation module of a large language model is adopted to generate a formalization statement St = pi (Q, Ht-1), and Q represents a natural language mathematical problem; carrying out self-verification Ct = pi (Q, Ht-1, St) by adopting a semantic self-verification module of a large language model; adding the current statement-evaluation pair (St, Ct) into a historical record to form Ht = Ht-1 U {(St, Ct)}; judging a termination condition: if the Ct represents that the semantics are consistent, outputting a formal statement Ans = St with consistent semantics, and terminating; otherwise, t is equal to t + 1, iteration is continued, and the steps S2-S4 are repeated. The reflection-self-correction mechanism endows the model with an iterative improvement capability, and semantic errors can be actively recognized and corrected. Experimental data show that on the basis of four standard benchmark tests, the semantic consistency is averagely improved by 17.2%, and the semantic consistency is improved by 30% on the basis of the most challenging AIME2025 data set.

An information processing apparatus acquires a user setting related to a feature of adversarial training and performs the adversarial training that trains a neural network by using training data including an adversarial sample and correct answer data indicating an original classification class, and the user setting, the adversarial training training the neural network that outputs a misclassification class in a case where the adversarial sample is input so as to output the original classification class in a case where the adversarial sample is input. The apparatus determines, by executing a format validation algorithm, that the adversarial sample does not exist within a predetermined range of noise of specific data in the neural network using a weighting factor obtained by the adversarial training.

The invention discloses a TSN time synchronization analysis method based on a formalization method. The TSN time synchronization analysis method comprises the following steps: step 1, knowing a TSN time synchronization mechanism according to a TSN time synchronization specification; 2, abstracting a formal description of TSN time synchronization by using a calendar automaton according to a mechanism described in a TSN time synchronization specification; 3, according to the formal description of TSN time synchronization, establishing a TSN time synchronization model which comprises a master clock model and a slave clock model; and 4, performing simulation and analysis by using the constructed TSN time synchronization model. The TSN time synchronization analysis method provided by the invention has relatively strong complex time sequence behavior expression capability, and the calendar automaton has natural support for multiple time granularities, periodic events and abnormal events, so that complex time sequence constraints in TSN time synchronization can be flexibly modeled; the limitation of a traditional timestamp and frequency compensation mechanism in expressing a complex scheduling strategy is overcome.

The application discloses a cloud-based scheduler conflict detection method based on a formal method, and comprises the following steps: step one, input and / or import affinity and / or anti-affinity rules, and convert the rules into a unified intermediate representation form of constraints, so as to shield the syntax and semantic differences of affinity and / or anti-affinity expressions between different schedulers; step two, convert the intermediate representation form obtained in step one into a format recognizable by a solver again, so as to convert the affinity and / or anti-affinity management problem into a formal satisfiability problem; and step three, check the satisfiability of the constraint set by the solver, and for the rules that do not satisfy the constraints, locate the source file of the conflict rules and the position of the rules in the source file according to metadata. The application further discloses a conflict detection system for realizing the conflict detection method, and application of the method or system.

The invention provides a structure-to-instance theorem automatic formalization method and device, and relates to the technical field of natural language process.The method comprises the steps that mathematical structure types of non-formalized problem description texts are recognized, and formalization files from structures to instances are generated for each mathematical structure type; performing type check on the generated code of the formalized file, and when the check is not passed, correcting the generated code according to a preset error correction strategy to obtain a skeleton file lacking theorem proof; calling a large model to generate a certification code of each theorem certification, and synthesizing the certification codes into the skeleton file to obtain a formalized code file; and translating the formalized code file back to obtain a formalized problem description text corresponding to the problem description text. The method is used for solving the problems that in the prior art, automatic formalization work lacks the structural understanding ability for mathematical knowledge, and the automatic formalization deduction process from the structure to the theorem is difficult to achieve.