An agent autonomous target decision system and method based on state space reasoning

Abstract

This invention discloses an intelligent agent autonomous goal decision-making system and method based on state-space reasoning, belonging to the field of artificial intelligence and cognitive computing technology. The system includes a processor, a memory, and a sensor interface. The processor executes program instructions in the memory to implement the following modules: a state-space storage module for storing a complete set of state types composed of six-dimensional binary codes; a state perception module for representing the agent's current state as a six-dimensional energy vector; a goal reasoning module for determining a set of candidate target states from a state transition rule base based on the current state code and context parameters, and calculating the optimal score for each candidate target state through a candidate evaluation function; a path planning module for planning a state transition path from the current state to the selected target state; and a preference learning module for modulating the goal reasoning results based on historical transition experience. This invention enables an intelligent agent to autonomously reason about target states and autonomously plan transition paths in dynamic environments, and is applicable to scenarios requiring autonomous state management, such as robots, autonomous driving, and intelligent decision-making systems.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and cognitive computing technology, specifically relating to an intelligent agent autonomous goal decision-making system and method based on state space reasoning, which can be used in application scenarios that require autonomous state management, such as robots, autonomous driving, and intelligent decision-making systems. Background Technology

[0002] Existing intelligent agent systems have the following fundamental limitations in state management and goal decision-making: (1) Passivity in goal setting. The goals of existing intelligent agent systems are usually specified by external instructions or preset tasks. The system itself lacks the ability to autonomously reason about the goal based on the current situation in a dynamic environment. When external instructions are missing or the environment changes, the system cannot autonomously determine "what state should it go to next" and can only maintain the current state or execute the preset default strategy.

[0003] (2) Lack of structured representation of causal relationships between states. Existing systems typically encode the relationships between states implicitly in a black-box manner (such as neural networks), lacking explicit and interpretable causal mapping. This makes it impossible for the system to perform structured goal reasoning in unknown situations, and also makes it impossible to explain "why this goal was chosen instead of that goal".

[0004] (3) The evaluation dimension of target selection is singular. Existing systems usually select targets based on only a single dimension (such as reward maximization or distance minimization), lacking a comprehensive evaluation of multiple dimensions such as state certainty, situational adaptability, and safety boundary. In complex dynamic environments, evaluation based on a single dimension often leads to suboptimal choices.

[0005] (4) Lack of learning target preferences from historical experience. Although existing systems can optimize strategies through reinforcement learning and other methods, they usually require a lot of trial and error and external reward signals. The system lacks a lightweight preference learning mechanism based on its own internal state changes, and cannot continuously optimize target selection without relying on external labels.

[0006] (5) Lack of multi-step path planning from the current state to the target state. After determining the target, existing systems usually jump directly to the target state or only plan a single-step transition, lacking the ability to plan multi-step evolution paths in a structured way.

[0007] Prior to this filing date, the applicant's earlier patent applications (such as CN202610506445.X, CN202610601934.3, etc.) have disclosed a state representation method based on a six-dimensional energy vector, a complete state space of sixty-four hexagrams, and an energy evolution mechanism. However, the aforementioned earlier applications mainly focus on "how the system perceives and converges to the target state," and have not yet provided complete technical solutions for "how the system autonomously infers the target state when there is a lack of external instructions, how to plan the evolution path from the current state to the target state, and how to learn target preferences from historical experience." Summary of the Invention

[0008] (a) Purpose of the invention The purpose of this invention is to provide an intelligent agent autonomous goal decision-making system and method based on state-space reasoning, so as to solve the technical problems in the prior art where intelligent agent systems lack autonomous goal reasoning ability, lack structured representation of causal relationships between states, have a single dimension for goal selection evaluation, lack a lightweight preference learning mechanism, and lack multi-step path planning ability.

[0009] Specific objectives include: (1) A structured target reasoning method based on state space is provided, which enables the agent to autonomously reason and determine the candidate target state based on the current state and context parameters without external instructions.

[0010] (2) Provides an explicit method for representing causal relationships between states, and realizes interpretable causal reasoning through a state transition rule base.

[0011] (3) A multi-dimensional candidate target evaluation method is provided, which integrates state certainty, situational fit, historical preference and safety boundary to select targets.

[0012] (4) A lightweight preference learning method based on changes in the internal state of the system is provided, which enables the system to continuously optimize the selection of objectives from historical experience without the need for external reward signals.

[0013] (5) Provides a multi-step path planning method from the current state to the target state, supporting the system to perform structured evolution path planning in the state space.

[0014] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: 1. System Architecture An intelligent agent autonomous goal decision-making system based on state-space reasoning includes a processor, a memory, and a sensor interface connected to the processor. The processor is configured to implement the following modules by executing program instructions stored in the memory: (1) State space storage module The state space storage module is used to store a set of basic state types consisting of six-dimensional binary codes.

[0015] The set of basic state types includes There are three basic state types. Each basic state type corresponds to a unique six-dimensional binary encoding vector. .

[0016] The six dimensions of the six-dimensional binary encoded vector correspond to the six basic aspects of the system's situation: foundation dimension (whether the system's foundation is stable), action dimension (whether proactive action is needed), information dimension (whether the information is clear and unambiguous), resource dimension (whether resources are sufficient), boundary dimension (whether the security boundary is clear), and environment dimension (whether the external environment is favorable). These six dimensions constitute the basic coordinate system of the system's state space.

[0017] The 64 basic state types mathematically constitute a complete state space—for any system state described by the six dimensions, a unique corresponding encoding vector can be found in the space.

[0018] The state space storage module also stores semantic description information for each basic state type, including state name, state attributes, and recommended action strategies corresponding to the state.

[0019] The state space storage module stores the basic state type set and semantic description information in the non-volatile storage area of ​​the memory, and loads them into the processor's running memory for use when the system starts.

[0020] (2) State awareness module The state perception module is used to acquire sensor input data through the sensor interface and represent the current state of the agent as a six-dimensional energy vector e=(e1,e2,e3,e4,e5,e6), where e_i∈[0,1].

[0021] The value of each dimension represents the system's tendency in that dimension: the closer the value is to 1, the more positive the state; the closer the value is to 0, the more negative the state. The initial state of the system is a neutral state where all dimensions are set to 0.5.

[0022] The sensor input data includes, but is not limited to: environmental perception data (camera, lidar, millimeter-wave radar), positioning data (GPS, inertial measurement unit), and system status data (battery level, temperature, operating mode).

[0023] The state perception module binarizes the six-dimensional energy vector at a preset threshold of 0.5 to obtain a six-dimensional binary encoded vector, and then matches it with the basic state type set to determine the current state code.

[0024] (3) Target reasoning module The target reasoning module is used to acquire sensor input data and system internal state data as context parameters. Based on the current state code and the context parameters, it determines a set of candidate target states from a preset state transition rule base and calculates the preferred score of each candidate target state through a candidate evaluation function.

[0025] The state transition rule base contains a causal relationship mapping between states, which defines the transition conditions and directions from a first state code to a second state code.

[0026] The causal mapping includes at least the following information: initial state code, target state code, transition conditions (including contextual parameter conditions), transition direction identifier (one-way / two-way), and transition weight.

[0027] The state transition rule base is a pre-configured, updatable rule base stored in the non-volatile storage area of ​​the memory, and supports the addition, modification, or deletion of rules through an external configuration file.

[0028] The target inference module performs multi-step inference with a configurable inference depth: - When the reasoning depth is one, the target reasoning module starts from the current state code, queries all causal mappings in the state transition rule base that start with the current state code, and adds the corresponding target state code to the candidate target state set. - When the inference depth is greater than one, the target inference module starts from the current state code and determines the candidate target state set after passing through one or more intermediate state codes. Specifically, for inference depth d (d>1), the target inference module takes the current state code as the starting point and performs d-step state transition inference: in the k-th step (1≤k≤d), taking the state code reached in the (k-1)-th step as the starting state, it queries all causal mappings in the state transition rule base that start with that state code, adds the corresponding target state code to the candidate set in the k-th step; and takes all state codes reached in the d-th step as the candidate target state set.

[0029] When the reasoning depth is greater than one, the target reasoning module scores the reasoning paths and selects the terminal state corresponding to the reasoning path with the highest score as the candidate target state. The score of the reasoning path is calculated based on the product or weighted sum of the transition weights of each state transition in the path.

[0030] The candidate evaluation function includes at least one of the following evaluation dimensions: - State certainty evaluation: calculated based on the Hamming distance between the current state code and the candidate target state code. A smaller Hamming distance indicates fewer changes are required to transition from the current state to the candidate target state, resulting in higher certainty. - Context fit assessment: Calculation of the correlation between context parameter vectors and candidate target state codes. Context parameters include, but are not limited to, task type, environmental state, resource availability, time constraints, and security level. - Historical preference assessment: Calculation based on preference modulation parameters generated by the preference learning module; - Security boundary assessment: Determined based on whether the candidate target state belongs to a preset set of conservative states. If the candidate target state belongs to the set of conservative states, the security boundary assessment score is higher.

[0031] The specific calculation method for the context fit evaluation is as follows: The context parameter vector is converted into a six-dimensional context vector; the dot product of the six-dimensional context vector and the candidate target state code is calculated; and this product is divided by the product of the context vector magnitude and the state vector magnitude to obtain the normalized correlation value. The specific formula is: A(candidate) = (V_context · V_candidate) / (||V_context|| × ||V_candidate||) Where V_context is the six-dimensional context vector obtained by transforming the context parameters, V_candidate is the candidate target state encoding vector, and · represents the vector dot product. When the context parameters and the candidate target state are highly matched, the correlation value approaches 1; when they do not match, the correlation value approaches 0.

[0032] The candidate evaluation function is a weighted sum of the evaluation dimensions: Score(candidate) = w1× D(candidate) + w1× A(candidate) + w1× P(candidate) + w1× S(candidate) Where D is the state certainty assessment value, A is the situation fit assessment value, P is the historical preference assessment value, S is the safety boundary assessment value, and w1, w2, w3, and w5 are the weights of each assessment dimension, and the weights are configurable parameters.

[0033] The target reasoning module selects the candidate target state with the highest optimization score as the target state.

[0034] (4) Path planning module The path planning module is used to plan a state transition path from the current state to the target state based on the current state code, the selected target state code, and the state transition rule base.

[0035] The state transition path is represented by a sequence of one or more intermediate state codes: Path = [S0, S1, S2, ..., Sn] Where S0 is the current state code, Sn is the target state code, and S1 to Sn-1 are intermediate state codes.

[0036] The state transition path planned by the path planning module satisfies the following constraints: - The Hamming distance between two adjacent state codes in the path is one (i.e., only the value of one dimension is changed at each step). - Each state transition in the path satisfies the transition conditions defined in the state transition rule base; - The length n of the state transition path is the minimum number of state transition steps required to go from the current state code to the target state code, or the number of steps that satisfy the preset path length constraint.

[0037] The path planning module uses a graph search algorithm (preferably A* algorithm or Dijkstra algorithm) to search for paths in the state space, with Hamming distance as the heuristic function and transition weights in the state transition rule base as edge costs.

[0038] When changes in the dynamic environment render the original path infeasible, the path planning module triggers path replanning, re-executing the path search from the current actual arrival state code.

[0039] (5) Preference Learning Module The preference learning module is used to record the starting state code, target state code, and transition result for each state transition, generate preference modulation parameters based on historical transition experience, and superimpose the preference modulation parameters into the calculation of the candidate evaluation function.

[0040] The transfer result includes the feedback signal of this transfer. The feedback signal is automatically generated by the system based on the direction of change of the statistical variance of the six-dimensional energy vector after the state transition relative to the statistical variance of the six-dimensional energy vector before the transition, without relying on any external labels or manual annotations. - When the statistical variance after the transfer is lower than the statistical variance before the transfer, it is judged as positive feedback ("good"), indicating that the transfer makes the system state more certain; - When the statistical variance after the transfer is higher than the statistical variance before the transfer, it is judged as negative feedback ("bad"), indicating that the transfer makes the system state more uncertain.

[0041] The statistical variance is defined as follows: U = (1 / 6)∑(e_i - ē)^2 Where ē is the mean of the six-dimensional energy vector, and the U value measures the degree of divergence between the dimensions of the system.

[0042] The preference learning module maintains a preference matrix M of size N×N (N=64), which is stored in the non-volatile storage area of ​​the memory. M[i,j] stores historical statistical information from state i to state j, including: number of transitions, number of positive feedbacks, number of negative feedbacks, and the most recent transition time.

[0043] The preference learning module calculates the preference modulation parameter P(candidate) based on the preference matrix: For the transition from the current state c to the candidate target state t: If the total number of transitions of M[c,t] is less than the preset threshold (preferably 3 times), then P(candidate) = 0; If the total number of transitions in M[c,t] is not less than a preset threshold, then: P(candidate) = (Number of positive feedbacks / Total number of transfers - 0.5) × 2 × R Where R is the preset preference weight range, preferably 0.2. The value range of P(candidate) is [-0.2, 0.2].

[0044] The preference learning module updates the preference matrix after each state transition: it stores the starting state code, target state code, and feedback signal (positive / negative) of this transition into the preference matrix; and it updates the total number of transitions and the number of positive / negative feedbacks for this transition path.

[0045] 2. Methodology and Flow A method for autonomous goal decision-making by an intelligent agent based on state-space reasoning includes the following steps: Step S1: State awareness.

[0046] The sensor input data is obtained through the sensor interface, and the current state of the agent is represented as a six-dimensional energy vector e=(e1,e2,e3,e4,e5,e6), where e_i∈[0,1].

[0047] The six-dimensional energy vector is binarized using a preset threshold of 0.5 to obtain a six-dimensional binary encoded vector, which is then matched with a set of basic state types to determine the current state code. The set of basic state types includes... There are 10 basic state types, and each type corresponds to a unique six-dimensional binary encoding vector.

[0048] Step S2: Target reasoning.

[0049] Sensor input data and system internal state data are acquired as context parameters. Based on the current state code and the context parameters, a preset state transition rule base is queried to determine a set of candidate target states.

[0050] The state transition rule base contains a causal relationship mapping between states, which defines the transition conditions and directions from a first state code to a second state code. The state transition rule base is stored in the non-volatile storage area of ​​the memory.

[0051] The optimal score of each candidate target state is calculated using the candidate evaluation function, and the candidate target state with the highest optimal score is selected as the target state.

[0052] The candidate evaluation function includes at least one of the following evaluation dimensions: state certainty evaluation, situation fit evaluation, historical preference evaluation, and safety boundary evaluation. The candidate evaluation function is a weighted sum of the evaluation dimensions, and the weights of each evaluation dimension are configurable parameters.

[0053] Step S3: Path planning.

[0054] Based on the current state code, the target state code, and the state transition rule base, a state transition path from the current state to the target state is planned.

[0055] The state transition path is represented by a sequence of one or more intermediate state codes. The Hamming distance between any two adjacent state codes in the path is one. The length of the state transition path is the minimum number of state transition steps required to reach the target state code from the current state code, or the number of steps that satisfy a preset path length constraint.

[0056] Step S4: State transition execution and preference learning.

[0057] Perform a state transition, record the starting state code, target state code, and transition result of this transition, and update the preference learning parameters.

[0058] The transfer result includes the feedback signal of this transfer. The feedback signal is automatically generated by the system based on the direction of change of the statistical variance of the six-dimensional energy vector after the state transfer relative to the statistical variance of the six-dimensional energy vector before the transfer: when the statistical variance after the transfer is lower than the statistical variance before the transfer, it is determined to be positive feedback; when the statistical variance after the transfer is higher than the statistical variance before the transfer, it is determined to be negative feedback.

[0059] The preference learning parameters are updated based on the proportion of positive feedback to total feedback.

[0060] (III) Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects: (1) The intelligent agent has achieved autonomous reasoning ability for the target state. Through the state transition rule base and candidate evaluation function, the present invention enables the intelligent agent to autonomously reason and determine the optimal target state based on the current state and context parameters in the absence of external instructions, thus eliminating the dependence on external instructions.

[0061] (2) It provides an explicit and interpretable representation of causal relationships between states. This invention explicitly encodes the causal relationship mapping between states through a state transition rule base, making the system's goal reasoning process interpretable—the system can trace "why this goal was chosen," rather than a black-box decision.

[0062] (3) Achieved comprehensive evaluation of candidate targets in multiple dimensions. This invention selects targets by integrating four dimensions: state certainty, situational fit, historical preference, and safety boundary. This avoids the one-sidedness of single-dimensional evaluation and improves the robustness and adaptability of target selection.

[0063] (4) A lightweight preference learning mechanism based on changes in the internal state of the system is provided. The preference learning module of this invention uses the system's own degree of certainty change (U value decrease / increase) as a feedback signal, without the need for external reward signals or manual annotation, and realizes fully self-supervised preference learning, enabling the system to continuously optimize target selection from historical experience.

[0064] (5) Multi-step path planning from the current state to the target state is realized. This invention supports the system to plan multi-step evolution paths from the current state to the target state through graph search in the state space by the path planning module, and supports path replanning in dynamic environments.

[0065] (6) Configurable inference depth. The target inference module of the present invention supports configurable inference depth, which can flexibly adjust the number of inference steps according to the complexity and real-time requirements of the application scenario, and achieve a balance between inference depth and computational efficiency.

[0066] (7) High versatility. This invention does not depend on specific application scenarios and can be used as a general autonomous goal decision module for various intelligent agent systems (robots, autonomous driving, intelligent decision-making systems, etc.).

[0067] (8) Flexible Deployment. This invention can be deployed in edge computing devices, including vehicle domain controllers, robot main control chips, or embedded intelligent processing units. The state transition rule base and preference matrix are stored in the non-volatile memory of the edge computing device and loaded into the processor's running memory for use when the system starts. Attached Figure Description

[0068] Figure 1 : Module architecture diagram of the intelligent agent autonomous goal decision-making system based on state space reasoning described in this invention.

[0069] Figure 2 A schematic diagram of a six-dimensional binary complete state space.

[0070] Figure 3 : Multi-step reasoning flowchart of the target reasoning module.

[0071] Figure 4 A diagram illustrating the multi-dimensional comprehensive scoring of candidate target evaluation.

[0072] Figure 5 : Schematic diagram of path search in the path planning module.

[0073] Figure 6 Flowchart of preference matrix update in preference learning module.

[0074] Figure 7 : The overall flowchart of the autonomous goal decision-making method described in this invention. Detailed Implementation

[0075] It should be understood that the following embodiments are used to illustrate the technical solutions of the present invention and not to limit them. Those skilled in the art can implement the complete technical solutions of the present invention based on the teachings of this specification.

[0076] The system can be deployed in an edge computing device, which includes an in-vehicle domain controller, a robot main control chip, or an embedded intelligent processing unit. The state transition rule base and preference matrix are stored in the non-volatile memory of the edge computing device and loaded into the processor's running memory when the system starts for use by the target inference module and preference learning module.

[0077] In a preferred embodiment of the present invention, at least a portion of the causal relationship mappings in the state transition rule base are configured based on traversal rules of the completeness of a six-dimensional binary vector space. This six-dimensional binary vector space contains... A non-repeating state encoding is used, with a Hamming distance of one between adjacent state encodings, forming a structured state transition network. This configuration leverages the completeness and symmetry of the state space, providing a deterministic and regularized source of state transition rules for this invention.

[0078] Example 1: System Initialization and State Space Construction This embodiment demonstrates the system initialization process and the construction of the state space.

[0079] Step 1: Define the state space.

[0080] System predefined There are six basic state types. Each state type corresponds to a unique six-dimensional binary encoding vector. The six dimensions are: foundation dimension, action dimension, information dimension, resource dimension, boundary dimension, and environment dimension.

[0081] The following are some examples of state type definitions: All positive state [1,1,1,1,1,1] Full-dimensional hyperactivity, maximum expansion Proactively promote and take positive action All negative state [0,0,0,0,0,0] Full-dimensional silence, maximum convergence Limit movement and conserve energy Stop observation state [0,0,1,0,0,1] Stop where you want to stop, suspend observation. Pause, observe, wait Humility and uprightness [0,0,1,0,0,0] Humility, integrity, restraint, and defensiveness Conservative strategy, defensive stance Open interactive state [1,1,1,0,1,1] Open interaction, joyful sharing Active interaction and information sharing Cautious travel mode [0,1,1,0,1,0] Proceed with caution and observe your surroundings. Proceed slowly and observe continuously Step 2: Initialize the state transition rule base.

[0082] The system pre-configures a state transition rule base, defining causal relationship mappings between states. These causal relationship mappings include at least the initial state code, the target state code, and the transition conditions.

[0083] An example partial causal mapping is as follows: - The state of "stop observing" [0,0,1,0,0,1] can be transitioned to the state of "cautious passage" [0,1,1,0,1,0] under the condition of "obstacle cleared"; - The "open interactive state" [1,1,1,0,1,1] can transition to the "fully positive state" [1,1,1,1,1,1] under the condition of "sufficient information"; - The state from "all positive" [1,1,1,1,1,1] can be transformed into "humble and positive" [0,0,1,0,0,0] under the condition of "resource scarcity"; - From the "cautious passage state" [0,1,1,0,1,0], you can transition to the "stop observing state" [0,0,1,0,0,1] under the condition of "obstacle detected".

[0084] The state transition rule base supports adding, modifying, or deleting rules through external configuration files.

[0085] Step 3: Initialize the preference matrix.

[0086] The system initializes the preference matrix M, which is 64×64 in size and stored in non-volatile memory, with all elements initialized to 0. Each element M[i,j] of the preference matrix stores the transition statistics from state i to state j, including the total number of transitions, the number of positive feedbacks, and the number of negative feedbacks.

[0087] Step 4: Initialize system state.

[0088] The system initializes the six-dimensional energy vector to a neutral state: e = [0.5, 0.5, 0.5, 0.5, 0.5, 0.5]. Through binarization matching, the current state is determined to be encoded as a completely negative state [0,0,0,0,0,0].

[0089] Example 2: Single-Step Target Reasoning This embodiment demonstrates the target reasoning process of the system when the reasoning depth is one.

[0090] Step 1: State awareness.

[0091] The system acquires sensor input data through the sensor interface, sensing the current six-dimensional energy vector e = [0.3, 0.7, 0.6, 0.4, 0.8, 0.2]. After binarization, the current state code S0 = [0, 1, 1, 0, 1, 0] is obtained, which matches the "cautious passage state".

[0092] Step 2: Obtain context parameters.

[0093] The system obtains current context parameters through the sensor interface: task type = "navigation", environment status = "obstacle detected ahead", resource availability = "sufficient", and security level = "medium".

[0094] Step 3: Target reasoning (reasoning depth = 1).

[0095] The target inference module starts with the current state code S0 = [0,1,1,0,1,0] and queries the state transition rule base to obtain all causal mappings starting with S0. Let the query results contain three candidate targets: - Candidate A: Transition to the "stop observation state" [0,0,1,0,0,1], with the transition condition being "obstacle detected"; - Candidate B: Transition to "humble and normal state" [0,0,1,0,0,0], with the transition condition being "resource scarcity"; - Candidate C: Transition to "open interactive state" [1,1,1,0,1,1], with the transition condition being "sufficient information".

[0096] Step 4: Candidate evaluation.

[0097] The candidate evaluation function calculates the optimal score for each candidate objective: Candidate A (stop observation state): - State determination assessment: The Hamming distance between S0 and [0,0,1,0,0,1] is 2, and the normalized score is 0.67; - Context fit assessment: "Obstacle detected" perfectly matches the current context "Obstacle detected ahead". Calculated using the correlation formula A(candidate) = (V_context · V_candidate) / (||V_context|| × ||V_candidate||), the score is 1.0. - Historical Preference Assessment: This transfer history is 0 times, and the score is 0; - Safety boundary assessment: The stopped observation state belongs to the set of conservative states, with a safety score of 1.0; - Overall score (all weighted at 0.25): (0.67+1.0+0+1.0) / 4 = 0.667.

[0098] Candidate B (Humble and Normal): - State certainty assessment: Hamming distance is 3, normalized score is 0.50; - Context fit assessment: "Resource shortage" does not match the current context of "resource abundance", with a correlation score of 0.2; - Historical preference assessment: 0; - Safety boundary assessment: The humble normality belongs to the set of conservative states, with a safety score of 1.0; - Overall score: (0.50+0.2+0+1.0) / 4 = 0.425.

[0099] Candidate C (Open Interactive State): - State certainty assessment: Hamming distance is 3, normalized score is 0.50; - Context fit assessment: "Information sufficient" partially matches the current context "resources sufficient", but the key feature of the current context is "obstacle detected", with a correlation score of 0.4; - Historical preference assessment: 0; - Security boundary assessment: The open interactive state does not belong to the set of conservative states, and the security score is 0.5; - Overall score: (0.50+0.4+0+0.5) / 4 = 0.350.

[0100] The target reasoning module selects the candidate A (stop observation state) with the highest optimization score as the target state.

[0101] Example 3: Multi-step target reasoning This embodiment demonstrates the target reasoning process of the system when the reasoning depth is greater than one.

[0102] The current system state code is S0 = [0,0,1,0,0,0] (humble normality), and the context parameters are: task type = "exploring unknown environment" and resource availability = "sufficient".

[0103] The target inference module performs multi-step inference with an inference depth of d=3: Step 1 Reasoning: Starting from S0, query the state transition rule base to obtain the set of states that can be reached in one step, S1 = {s11, s12, s13}.

[0104] Step 2 Reasoning: Starting from each state in S1, query the state transition rule base to obtain the two-step reachable state set S2 = {S21, S22, S23, S24}.

[0105] Step 3 Reasoning: Starting from each state in S2, query the state transition rule base to obtain the three-step reachable state set S3={S31, S32, S33, S34, S35}.

[0106] For each state in S3, calculate its corresponding inference path score. The inference path score is calculated based on the product of the transition weights of each state transition in the path.

[0107] Three candidate reasoning paths and their scores are as follows: - Path 1: S0→ S11→ S21→ S31, Path score 0.85 × 0.90 × 0.95 = 0.727; - Path 2: S0→ S11→ S22→ S32, Path score 0.85 × 0.70 × 0.88 = 0.524; - Path 3: S0→ S12→ S23→ S33, path score 0.60 × 0.85 × 0.92 = 0.469.

[0108] The target reasoning module selects the terminal state s corresponding to path 1, which has the highest path score. 31 As a candidate target state, add it to the candidate target state set.

[0109] Then, the optimal score for each candidate objective is calculated using the candidate evaluation function, and the objective with the highest score is selected as the final objective.

[0110] Example 4: Path Planning This embodiment demonstrates the path planning process of the system from the current state to the target state.

[0111] The current state code is S0 = [0,0,1,0,0,1] (stop observation state), and the target state code is s_target = [1,1,1,1,1,1] (all positive state).

[0112] The path planning module searches for a path in the state space, starting from S0 and ending at s_target.

[0113] During the search process, the Hamming distance between two adjacent state codes in the path is one. The Hamming distance is used as the heuristic function h(s) = HammingDistance(s, s_target), and the transition weights in the state transition rule base are used as the edge costs.

[0114] The optimal path obtained by the path planning module is: s0[0,0,1,0,0,1] → [0,1,1,0,0,1] → [0,1,1,1,0,1] → [0,1,1,1,1,1] →[1,1,1,1,1,1] (target state).

[0115] Path length: 4 steps. Each step changes only one dimension, and each step satisfies the transition conditions defined in the state transition rule base.

[0116] When changes in the environment (such as "resource availability" changing from "sufficient" to "scarce") render the original path infeasible, the path planning module re-executes the path search starting from the current actual arrival state code to generate a new feasible path.

[0117] Example 5: Preference Learning and Modulation This embodiment demonstrates how the system uses historical experience to modulate the preference for target selection.

[0118] In the initial preference matrix M of the system, the historical record from state A (open interactive state, encoded [1,1,1,0,1,1]) to state B (all positive state, encoded [1,1,1,1,1,1]) is as follows: a total of 10 transitions, of which the U value decreased after 8 transitions (positive feedback), and the U value increased after 2 transitions (negative feedback), with a positive feedback rate of 80%.

[0119] The historical record of transitioning from state A to state C (humble normality, encoded as [0,0,1,0,0,0]) is as follows: a total of 3 transitions, with the U value decreasing after 1 transition (positive feedback) and the U value increasing after 2 transitions (negative feedback), resulting in a positive feedback rate of 33%.

[0120] When the system calculates the candidate target again from state A: Preference modulation parameter for candidate target B (all positive state): P(B) = (0.8 - 0.5) × 2 × 0.2 = 0.12.

[0121] Preference modulation parameter for candidate target C (humble normality): P(C) = (0.33 - 0.5) × 2 × 0.2 = -0.068.

[0122] The system superimposes the preference modulation parameters into the calculation of the candidate evaluation function, so that the candidate target B can obtain a higher optimization score, thus giving priority to the "all positive state" as the target.

[0123] If the U value decreases after this transition, the system automatically determines it as positive feedback and updates the preference matrix M[A,B]. The total number of transitions is 11, the number of positive feedbacks is 9, and the positive feedback rate of this path increases to 81.8%.

[0124] Example 6: Application Example with Autonomous Systems This embodiment demonstrates the application of the autonomous goal decision-making method of the present invention in an autonomous system.

[0125] In the robot scenario, the robot is currently in a "task execution" state, and the sensors have detected that "the battery level is below 20%".

[0126] Step S1: State perception. The system acquires sensor input data through the sensor interface, represents the current state as a six-dimensional energy vector, and after binarization, matches the current state encoding as "execution state" [1,1,1,1,0,1].

[0127] Step S2: Target Reasoning. The system acquires contextual parameters (battery level = 15%, task urgency = medium) through the sensor interface, queries the state transition rule base, and determines the candidate target set, including "charging state" [0,0,1,0,0,0] and "task completed state" [1,1,1,0,1,1]. The candidate evaluation function comprehensively considers state certainty, contextual fit ("low battery" and "charging state" are highly matched, with a correlation value of 0.95), historical preference (the positive feedback rate from "execution state" to "charging state" is 85% in history), and safety boundary ("charging state" belongs to the conservative state set), and selects "charging state" as the target.

[0128] Step S3: Path planning. The path planning module plans the transition path from the "execution state" to the "charging state": [1,1,1,1,0,1] → [0,1,1,1,0,1] → [0,0,1,1,0,1] → [0,0,1,0,0,1] → [0,0,1,0,0,1] → [0,0,1,0,0,0].

[0129] Step S4: Execution and Learning. The robot performs a state transition and navigates to the charging station to complete charging. After the transition is complete, the U-value decreases, which the system automatically interprets as positive feedback, and updates the preference matrix.

[0130] The autonomous goal decision-making method of the present invention serves as the core decision-making module of an autonomous system, enabling the system to autonomously reason about the goal state, autonomously plan the transfer path, and continuously optimize decisions based on historical experience in the absence of external instructions.

Claims

1. An agent-based autonomous goal decision-making system based on state-space reasoning, characterized in that, The processor includes a processor, a memory, and a sensor interface connected to the processor. The processor is configured to implement the following modules by executing program instructions stored in the memory: The state space storage module is used to store a set of basic state types composed of six-dimensional binary codes, the set of basic state types including There are three basic state types, each corresponding to a unique six-dimensional binary encoding vector. ; The state perception module is used to acquire sensor input data through the sensor interface, represent the current state of the agent as a six-dimensional energy vector e=(e1,e2,e3,e4,e5,e6), where e_i∈[0,1], and determine the current state code by binarizing the six-dimensional energy vector and matching it with the set of basic state types. The target reasoning module is used to acquire sensor input data and system internal state data as context parameters, determine a set of candidate target states from a preset state transition rule base based on the current state code and the context parameters, and calculate the preferred score of each candidate target state through a candidate evaluation function. The path planning module is used to plan a state transition path from the current state to the target state based on the current state code, the selected target state code, and the state transition rule base. The state transition path is represented by a sequence of one or more intermediate state codes. The preference learning module is used to record the starting state code, target state code, and transition result for each state transition, generate preference modulation parameters based on historical transition experience, and superimpose the preference modulation parameters into the calculation of the candidate evaluation function.

2. The system according to claim 1, characterized in that, The state transition rule base contains a causal relationship mapping between states, which defines the transition conditions and direction from a first state code to a second state code; the candidate evaluation function includes at least one of the following evaluation dimensions: state certainty evaluation, context fit evaluation, historical preference evaluation, and safety boundary evaluation; the state transition rule base is a pre-configured, updatable rule base stored in the non-volatile storage area of ​​the memory, and supports adding, modifying, or deleting rules through an external configuration file.

3. The system according to claim 2, characterized in that, The state certainty assessment is based on the Hamming distance between the current state code and the candidate target state code; the context fit assessment is based on the correlation between the context parameter vector and the candidate target state code; the historical preference assessment is based on the preference modulation parameters generated by the preference learning module; and the security boundary assessment is based on whether the candidate target state belongs to a preset conservative state set.

4. The system according to claim 1, characterized in that, The target reasoning module performs multi-step reasoning with a configurable reasoning depth: when the reasoning depth is one, the target reasoning module directly determines the candidate target state set from the current state code; when the reasoning depth is greater than one, the target reasoning module starts from the current state code and determines the candidate target state set after passing through one or more intermediate state codes.

5. The system according to claim 4, characterized in that, When the reasoning depth is greater than one, the target reasoning module scores the reasoning path and selects the terminal state corresponding to the reasoning path with the highest score as the candidate target state. The score of the reasoning path is calculated based on the product or weighted sum of the transition weights of each state transition in the path.

6. The system according to claim 1, characterized in that, In the preference learning module, the transition result includes the feedback signal of this transition; the feedback signal is automatically generated by the system based on the direction of change of the statistical variance of the six-dimensional energy vector after the state transition relative to the statistical variance of the six-dimensional energy vector before the transition: when the statistical variance after the transition is lower than the statistical variance before the transition, it is determined to be positive feedback; when the statistical variance after the transition is higher than the statistical variance before the transition, it is determined to be negative feedback; the preference modulation parameter is calculated based on the proportion of positive feedback times to the total number of feedback times.

7. The system according to claim 1, characterized in that, The state transition path planned by the path planning module satisfies the following constraints: the Hamming distance between two adjacent state codes in the path is one; the length of the state transition path is the minimum number of state transition steps required from the current state code to the target state code.

8. The system according to claim 1, characterized in that, The six dimensions of the six-dimensional binary encoding vector correspond to the following dimensions: foundation dimension, action dimension, information dimension, resource dimension, boundary dimension, and environment dimension.

9. A method for autonomous goal decision-making by an intelligent agent based on state-space reasoning, characterized in that, The method is performed by the system according to any one of claims 1 to 8, and includes the following steps: Step S1: Obtain sensor input data through the sensor interface and represent the current state of the agent as a six-dimensional energy vector e=(e1,e2,e3,e4,e5,e6), where e_i∈[0,1]; Step S2: Binarize the six-dimensional energy vector according to a preset threshold to obtain a six-dimensional binary encoded vector, and match it with the basic state type set to determine the current state code; the basic state type set includes There are 10 basic state types, each type corresponding to a unique six-dimensional binary encoding vector; Step S3: Obtain sensor input data and system internal state data as context parameters. Based on the current state code and the context parameters, query the preset state transition rule base to determine the candidate target state set. Step S4: Calculate the optimization score of each candidate target state using the candidate evaluation function, and select the candidate target state with the highest optimization score as the target state; Step S5: Based on the current state code, the target state code, and the state transition rule base, plan a state transition path from the current state to the target state. The state transition path is represented by a sequence of one or more intermediate state codes. Step S6: Perform state transition, record the starting state code, target state code, and transition result of this transition, and update the preference learning parameters.

10. The method according to claim 9, characterized in that, The causal mapping in the state transition rule base is configured using a traversal rule based on the completeness of the six-dimensional binary code space. This traversal rule corresponds to the non-repeating exhaustive encoding of sixty-four basic state types.

Images