A DECISION SUPPORT SYSTEM AND METHOD FOR AUTONOMOUS INTENTION ESTIMATION USING DYNAMIC UNCERTAINTY AND HISTORY MEMORY INTEGRATION IN A MULTIDIMENSIONAL STATE SPACE.

Abstract

The invention is a probabilistic decision support method and system that predicts the intentions of autonomous actors with high accuracy in complex and noisy environments. The method combines instantaneous state data, past action sequences, and environmental uncertainty into a single state vector, and calculates the most likely intention class and confidence coefficient using a special cognitive scoring function and MAP estimation. This reduces the rate of erroneous decisions in crisis situations with high uncertainty.

Description

1 TARIFF DYNAMIC UNCERTAINTY AND PAST MEMORY IN MULTIDIMENSIONAL STATE SPACE DECISION SUPPORT SYSTEM THAT PERFORMS AUTONOMOUS INTENTION PREDICTION THROUGH INTEGRATION AND METHOD TECHNICAL AREA 5 The invention is relevant to autonomous systems, decision support mechanisms, and defense technologies. probabilistically predicting the future intentions of actors and objects under complex / dynamic environmental conditions. It relates to a decision support system and method that predicts. The invention is particularly relevant to artificial intelligence. It falls within the fields of recursive neural networks and probabilistic state-space modeling. PRIOR TECHNIQUE (STATE BEFORE INVENTION) 10 Traditional decision support and intention prediction systems often only provide the target actors with immediate insights. It takes into account physical data such as position and speed (X, T). However, it is highly noisy, chaotic, and In dynamic environments, predictions based solely on physical data lead to high error rates and incorrect decisions. This leads to its being given. The deep learning models that are part of the current state of the art involve complex neural networks. Because they rush through it, it results in high transaction costs (computational complexity) and transaction delays. (latency) is the cause. This situation is problematic in autonomous systems that need to operate in real time. It slows down the decision-making process. In addition, current AI models deal with issues such as sensor noise or jamming in the environment. dynamically account for the "uncertainty level" arising from the factors. It is unable to do so. As a result, existing systems fail during times of crisis or high uncertainty. Overconfident individuals create flawed action plans, leading to disastrous decisions. It is possible. THE TECHNICAL PROBLEM THAT THE INVENTION SOLVED The purpose of the invention is to analyze not only instantaneous physical data but also context (C), past memory (H), targets (W), 25 Behavioral / emotional state (E) and instantaneous uncertainty in the system (U) in a single d-dimensional state. By combining and compressing the vector (𝑆𝑡), the processing load is reduced, thereby decreasing decision delay. The goal is to minimize it. However, the invention integrates dynamic risk and uncertainty into a cognitive scoring function. the term punishment (𝑢𝑡 Through (Tr), the probability of risky potential targets is 30 during times of high uncertainty. to automatically lower their scores and prevent overconfident, erroneous decisions in crisis situations a method and system that estimates the most reliable intent class (𝐼𝑡) and confidence score (conf𝑡) by passing through It aims to present. DESCRIPTION OF THE FIGURES FIGURE 1: Location (110), context (120) and emotional state (130) received from sensors 35 Showing the transformation of the data into the state vector (100) 𝑆𝑡 in the state space unit. It is system architecture. 2 FIGURE 2: Data ht from memory unit (200) and t from uncertainty unit (300). The data is processed in the cognitive scoring engine (400) and Softmax (500) and MAP decision unit It is an architectural diagram showing the conversion of (600) to the most likely intention (𝐼𝑡). FIGURE 3: Data collection (100), memory update (200), score calculation (400), probability 5 steps, including estimation (500), choice of intention (600) and action planning (700). It is a process flowchart. ANNOUNCEMENT OF THE INVENTION The invention describes a method that processes noisy data from sensors to generate autonomous action decisions. It consists of the following steps: 1. Creation of the State Space and Vector (𝑆𝑡): From sensor units and data 10 The information gathered from the sources is transformed into a d-dimensional state vector 𝑆𝑡: 𝑆𝑡 = [𝑥𝑡, 𝜏𝑡 , 𝑐𝑡 , ℎ𝑡 , 𝑔𝑡 , 𝑒𝑡 , 𝑢𝑡] ∈ ℝ d 𝑑 = 3 + 1 + 𝑑𝑐 + 𝑑ℎ + 𝑑𝑔 + 𝑑𝑒 + 𝑑𝑢 Here 𝑥𝑡 ∈ ℝ 3 location data, 𝜏𝑡 ∈ ℝ time, 𝑐𝑡 ∈ ℝ 𝑑𝑐 context, ℎ𝑡 ∈ ℝ dℎ past summary memory, 𝑔𝑡 ∈ ℝ 𝑑𝑔 is the target / requirement component, 𝑒𝑡 ∈ ℝ 𝑑𝑒 behavioral / emotional state and 𝑢𝑡 ∈ ℝ d dynamic 15 It represents the uncertainty component. 2. Historical Memory Encoder (ht): Past state (s), action (a), and The result / reward (𝑟) sequences are compressed via a Gated Recurrent Unit (GRU) cell to summarize ℎ𝑡. The vector is generated: ℎ𝑡 = GRU([𝑆𝑡−𝑘, 𝑎𝑡−𝑘, 𝑟𝑡−𝑘], … , [𝑆𝑡−1, 𝑎𝑡−1, ​​𝑟𝑡−1]) 20 3. Uncertainty Model (𝑢𝑡): Uncertainty in the environment or measurement is represented by the variance matrix we learned. It is modeled via a Gaussian distribution: 𝑢𝑡 ∼ 𝒩(𝜇𝑢, Σ𝑢), Σ𝑢 = diag(𝜎1 2, 𝜎2 2, … , 𝜎𝑑𝑢 2 ) 4. Cognitive Scoring Function (s𝑠𝑖(𝑆𝑡)): The system calculates the following for each candidate intention class (𝐼𝑖 ∈ ℐ): It runs the scoring equation: 25 𝑠𝑖(𝑆𝑡) = 𝑤0 +𝑤 𝑇𝜙(𝑆𝑡, 𝐼𝑖) + 𝑏𝑖 + 𝛾 ⋅ sim(ℎ𝑡,𝑚𝑖) + 𝛼 ⋅ 𝑢𝑡 𝑇𝑟𝑖 − 𝜆𝑅(𝐼𝑖) Here: • 𝜙(𝑆𝑡, 𝐼𝑖): This is the state-intent interaction feature transformation (MLP layer). • sim(ℎ𝑡,𝑚𝑖): This is the cosine similarity between the past summary ℎ𝑡 and the intention prototype 𝑚𝑖. • 𝑢𝑡 𝑇𝑟𝑖: This is the fit coefficient between the immediate uncertainty aspect and the intention risk profile (high uncertainty indicates a risky 30). (Intentions are punished). • R(I): Intent complexity regularization term (L1 / L2 norm). 3 5. Probability Distribution and Maximum A Posteriori (MAP) Estimation: Calculated scores are from Softmax. By subjecting it to the layer, the probability distribution (𝑃(𝐼𝑖 ∣ 𝑆𝑡)) is obtained. The highest probability intention (𝐼𝑡) and The confidence score (conft) is selected: 𝑃(𝐼𝑖 ∣ 𝑆𝑡) = exp(s<binary data, 4 bytes>(s<binary data, 4 bytes)) ∑ exp (sg(st)) N 𝑗=1 𝐼𝑡 = argmax 𝐼𝑖∈ℐ 𝑃(𝐼𝑖 ∣ 𝑆𝑡), conf𝑡 = 𝑃(𝐼𝑡 ∣ 𝑆𝑡) 5 6. Decision and Action Planning (𝑎𝑡): Based on the obtained 𝐼𝑡 intention estimation and conf𝑡 confidence coefficient. The autonomous system implements the relevant physical or digital counter-action plan (𝑎𝑡).

Claims

4 REQUESTS 1. Probabilistic intent via dynamic state space in an autonomous system or decision support mechanism. It is a method that makes predictions, and its characteristic feature is; a) Parameters received from sensor units and data sources; position (𝑋), time (𝑇), context 5 d- includes the components of (C), past memory (H), goals (H), behavioral state (E), and uncertainty (U). The step of converting it into a three-dimensional 𝑆𝑡 state vector, b) Past state, action, and outcome sequences through a recursive neural network unit (RNN). The step of creating a ℎt summary memory vector by processing, c) The generated 𝑆𝑡 state vector and ℎ𝑡 memory vector; instantaneous uncertainty (𝑢𝑡) and risk coefficient (𝑟𝑖) 10 The step of subjecting it to a scoring function 𝑠𝑖(𝑆𝑡) that includes the product, d) Converting the calculated scores into a probability distribution (P(I ∣ St)) using the Softmax activation function and According to the Maximum A Posteriori (MAP) principle, the most likely intention class (𝐼𝑡) and the confidence score (conf𝑡) the step of estimation, e) Depending on the predicted intent class (It), the autonomous action plan (at) is implemented by the implementing units. 15 It includes the steps to be implemented.

2. A method that conforms to Claim 1, and its characteristic is that the uncertainty term of the aforementioned scoring function 𝑠𝑖(𝑆𝑡) (𝑢𝑡 (by means of) reducing the probability scores of risky potential candidates during periods of high uncertainty It is the act of punishing.

3. An autonomous decision support system that runs the method corresponding to claim 1 or 2, and whose feature is data collection 20 a unit that generates physical or electronic countermeasures based on the processor unit, memory unit, and the anticipated intent (It). It includes an action execution unit.

Images