Railway signal scheduling optimization method and system based on AlphaEvove algorithm
By combining the AlphaEvolve algorithm and self-supervised learning, the railway signal dispatching system has achieved efficient collaborative operation and real-time response, solving the problems of insufficient intelligence and low efficiency of manual intervention in traditional dispatching systems, and improving the adaptability and response speed of the dispatching system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 卡斯柯信号(成都)有限公司
- Filing Date
- 2026-01-09
- Publication Date
- 2026-05-05
AI Technical Summary
Existing railway signal dispatching systems struggle to achieve efficient and coordinated operation when faced with complex and ever-changing railway operation scenarios, and lack the ability to respond to emergencies in real time. Traditional dispatching systems rely on manual intervention, resulting in low efficiency and insufficient intelligence.
A railway signal scheduling optimization method based on the AlphaEvolve algorithm is adopted, which combines Monte Carlo Tree Search (MCTS), policy-value network and self-supervised learning mechanism. A track state representation model is constructed through graph neural network, and an evolutionary strategy is introduced for scheduling optimization to generate an efficient scheduling scheme.
It has improved the learning and generalization capabilities of the railway dispatching system, reduced manual intervention, enhanced dispatching quality and response speed, adapted to changing railway operation scenarios, and reduced conflicts and delays.
Smart Images
Figure CN121973833A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of railway signaling technology, and more specifically to a railway signaling optimization method and system based on the AlphaEvolve algorithm. Background Technology
[0002] Railway signaling is a core component of the railway transportation dispatching system. It is primarily responsible for real-time monitoring, instruction issuance, and coordinated control of train operations through signaling equipment and communication technology, ensuring that trains run safely and efficiently according to the timetable. Current railway signaling technology suffers from the following problems: 1. Traditional railway signal dispatching systems face challenges of complexity and real-time performance. With increasing train density and more complex railway networks, railway signaling and dispatching systems must achieve efficient collaborative operation of multiple trains while ensuring safety. However, traditional dispatching systems rely heavily on rule-driven or static chart methods, making it difficult to respond in real time to emergencies such as train delays, temporary timetable adjustments, or line occupancy conflicts. Manual intervention by dispatchers is inefficient and prone to errors, severely hindering the improvement of railway transport capacity and the development of intelligent dispatching.
[0003] 2. Existing intelligent algorithms have limitations in their application in the railway field. In recent years, heuristic algorithms such as genetic algorithms, ant colony algorithms, and classic Monte Carlo Tree Search (MCTS) have been introduced into solving scheduling problems. However, these methods are mostly used for manufacturing systems or computational task scheduling, and have significant shortcomings in terms of real-time performance, safety, and topological constraint modeling of railway signaling systems. Furthermore, the lack of efficient state modeling methods and policy generalization mechanisms makes these methods difficult to adapt to complex and ever-changing railway operation scenarios, and unable to meet the needs of large-scale, multi-objective, and high-frequency scheduling decisions.
[0004] 3. The integration of deep learning and reinforcement learning technologies has not yet been fully adapted to the characteristics of railway operations. While deep reinforcement learning frameworks like AlphaGo have demonstrated exceptional decision-making and optimization capabilities, directly applying them to railway scheduling scenarios is impractical. This is primarily because railway scheduling problems are characterized by multiple objectives (safety, efficiency), strong constraints (signal sections, passing rules), and high real-time requirements. This necessitates that scheduling algorithms not only possess strong learning capabilities but also incorporate an understanding of railway business logic and graph structure modeling abilities. Currently, there is a lack of a scheduling optimization system that integrates deep learning, evolutionary optimization, and self-supervised state understanding to meet these requirements. Summary of the Invention
[0005] To overcome the shortcomings of the existing technology, this invention discloses a railway signal scheduling optimization method and system based on the AlphaEvolve algorithm. The AlphaEvolve algorithm (composed of Monte Carlo Tree Search (MCTS), policy-value network, evolutionary policy and self-supervised learning mechanism) is integrated to solve the problems of frequent train path conflicts, delayed scheduling response and low system intelligence.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a railway signal scheduling optimization method based on the AlphaEvolve algorithm, comprising the following steps: I. Orbit State Initialization S1. Construct the track state diagram and initialize the state diagram; Preferably, step S1 includes: obtaining the current railway operation diagram, train position, and equipment status, and using the obtained data to construct a track topology diagram, generate a track status diagram, and initialize the scheduling start status node in the status diagram as the root node of the Monte Carlo tree search.
[0007] II. Model Network Construction S2. Construct an orbital state representation model, a policy network, and a value network; wherein, the orbital state representation model is constructed based on a graph neural network and is used to encode the state graph and extract key orbital state features; the policy network is used to generate an initial scheduling probability distribution; and the value network is used to estimate the potential scheduling quality of the current state. In this invention, the "initial scheduling probability distribution" is mainly used in the "Monte Carlo Tree Search (MCTS) selection phase" of step S3. The "policy network output" of step S3 is precisely the "initial scheduling probability distribution" generated by the policy network in S2 based on the key track state characteristics.
[0008] The scheduling quality is related to the "simulation phase" and "feedback phase" of the S3 step. Specifically, their relationship and function are as follows: 1. Serving as the final evaluator during the "simulation phase" (instead of a full simulation): Traditional MCTS simulation: Starting from the newly expanded nodes, the simulation proceeds step by step through random or simple rules to deduce the "scheduling actions" until all trains reach the destination or a conflict is triggered, and a final result (such as total delay time) is obtained.
[0009] AlphaEvolve's improved simulation: When MCTS reaches the "simulation phase," it no longer performs a complete simulation. Instead, it inputs the "track state" corresponding to the newly expanded node into the value network already trained in S2. The value network immediately outputs an estimate V (i.e., "potential scheduling quality") for that state. This V value is a quick prediction of "how good the scheduling result can be obtained from this state."
[0010] Impact: This significantly reduces the computational cost of a single search. A single forward propagation of a neural network is much faster than a complete, multi-step scheduling deduction. This allows the algorithm to perform more search iterations in the same amount of time, thereby exploring a wider decision space.
[0011] 2. As a value signal in the "backhaul phase" for reverse propagation: The estimated value V (i.e., scheduling quality prediction) obtained in the “simulation phase” will be propagated backward from the leaf node to the root node along the MCTS search path in the “backhaul phase” to update the statistical information (mainly the action value Q) of all nodes on the path.
[0012] Function: This value signal is the core basis for all nodes to update their "good" or "bad" judgments. An action that leads to a high valuation V (high-quality scheduling prediction) will increase the Q value of the corresponding node, making it more likely to be selected in future searches. This completes the real-time online guidance of the search tree by the value network.
[0013] III. Monte Carlo Tree Search (MCTS) Process S3. Construct a Monte Carlo tree and use the selection, expansion, simulation, and backpropagation stages of the Monte Carlo tree search to perform the search. During the search process, for each state node in the tree, extract its key orbital state features using the orbital state representation model, and evaluate the state node through the policy network and value network to guide the search. After the search, generate a preliminary scheduling policy and use the state transition data generated during the simulation process to train the orbital state representation model in a self-supervised manner. Preferably, in step S3, the Monte Carlo tree search includes: Selection Phase: Using the initialized track state diagram as the root node, construct a search tree. Starting from the root node of the search tree, select the optimal scheduling action path among the child nodes according to the PUCT formula, and integrate the policy network output and the number of visits. Expansion phase: If the selected child node is not fully expanded, a new child node is added to represent the new scheduling decision; Simulation phase: A complete scheduling simulation is performed starting from this new child node; the scheduling simulation includes simulating trains passing through paths in sequence and determining conflicts / successes; Backhaul phase: The simulation results are backhauled to the nodes on the path to update the node estimates; wherein, the simulation results include whether there is a conflict, path cost, and delay cost.
[0014] IV. Evolutionary Strategy-Guided Sample Generation S4. Introduce multiple candidate scheduling sequences from the initial scheduling strategy as a population, and use an evolutionary strategy to generate new scheduling samples. Based on path cost, screen the new scheduling samples and update the strategy network. The screened new scheduling samples and the initial scheduling strategy form candidate samples. V. Output the optimal scheduling result S5. Select the scheduling scheme with the fewest conflicts and the least delay from the candidate samples as the optimal scheduling scheme and output it.
[0015] Preferably, in step S5, the output includes: train running path, passing time point table, and platform / turnout occupancy sequence.
[0016] Preferably, step S5 includes: firstly, selecting a subset of feasible schemes with fewer than the safety threshold from all candidate samples, and then selecting the scheme with the minimum total delay time from this subset as the optimal scheduling scheme, and outputting an executable scheduling instruction set.
[0017] I. Monte Carlo Tree Search 1. Selection Phase Preferably, the selection phase of the Monte Carlo tree search includes: The selection phase process includes: starting from the root node of the search tree, selecting the optimal scheduling action path among the child nodes according to the PUCT strategy, until a leaf node or a node that has not yet been expanded is reached; The strategies applied to railway scheduling during the selection phase include: Node representation: Each node represents a train scheduling status; Action selection considers: the current set of trains available for dispatch, the priority of each train, and the availability of the current signal section; The initial scheduling probability distribution and the number of visits provided by the policy network are used for balancing, including:
[0018] In the formula, A comprehensive evaluation of the action; For the value of the action; To explore and utilize the equilibrium constant; This represents the initial scheduling probability distribution; This represents the number of times the parent node has been visited. This represents the number of times the action is accessed.
[0019] 2. Expansion Phase Preferably, the extended phase of the Monte Carlo tree search includes: The expansion phase process description includes: if the currently selected node is not fully expanded, add a child node representing the new scheduling action to it; The rules applied to railway scheduling during the extended phase include: For a scheduling state, the actions that can be extended include: arranging a train to enter a certain signal section, assigning a train the next running path, and adjusting the departure interval between trains; Business constraints need to be considered: preventing two trains from occupying the same block section at the same time, and hard constraints on speed limits and turnout status.
[0020] 3. Simulation Phase Preferably, the simulation phase of Monte Carlo tree search includes: The simulation phase process description includes: starting from the newly expanded child nodes, performing a complete scheduling simulation based on the current state until the end state, and using the value network to evaluate the scheduling quality; The simulation process applied to railway dispatching includes: Use a policy network to guide the simulation direction, and add a certain degree of randomness to maintain diversity; The simulation generates the scheduling trajectory, including: train running path, passing time of each signal, and occupied section record; Output simulated evaluation values: Negative factors: number of conflicts, delay time, and route deviation; Positive factors: overall traffic efficiency, and good alignment with traffic schedules; A simulated return value is generated by combining all the data.
[0021] 4. Backhaul stage Preferably, the backhaul phase of a Monte Carlo tree search includes: The backhaul phase process includes: uploading the reward values obtained in the simulation phase along the path to all traversed nodes to update the average value Q and access count N of each node. The backhaul formula is:
[0022]
[0023] in, For the first The number of visits to each node; For the first The average value of each node; This is an estimate of the state value; The meaning of the return transmission phase in railway dispatching includes: A higher node access frequency indicates a more optimal scheduling path; A high Q value indicates that the corresponding scheduling action has a good historical performance; Backhauls enhance high-quality scheduling paths and suppress inefficient or insecure paths.
[0024] II. Evolutionary Strategies Preferably, step S4 includes: S41. Initialize the candidate individual population: Obtain several high-scoring scheduling sequences from Monte Carlo tree search as the population. Each individual in the population contains: the running path of each train, the departure time plan, and the signal occupancy plan. S42. Individual mutation to generate new samples: Use perturbation operators to perform individual mutation to generate new scheduled samples; Preferably, in step S42, the perturbation operator includes: Time disturbance: Minor adjustments to train departure / arrival times; Path replacement: Select an equivalent branch from the feasible paths; Sequence rearrangement: Adjusting the dispatching order of two trains; Sectional exchange: exchanging the train arrival order at a certain junction.
[0025] S43. Fitness function calculation: Using the fitness function, calculate the score of scheduling new samples under multi-objective conditions; Preferably, in step S43, the fitness function is:
[0026] in, For new samples The score, This represents the number of signal collisions. Total train delay time; To determine whether the original operating schedule is met; Weight passengers according to their priority. , , , This is an adjustable coefficient that reflects the scheduling strategy tendency.
[0027] S44. Selection and Feedback of Winning Individuals: Select the top k% of the new scheduling samples as winners, and use the winners to update the policy network, to expand the new actions of nodes on the Monte Carlo tree search tree, and to serve as the initial individuals for the next round of evolution.
[0028] Secondly, based on the above railway signaling scheduling optimization method, this invention also provides a railway signaling scheduling optimization system based on the AlphaEvolve algorithm, comprising: The track topology modeling module uses a graph structure to represent the railway signaling system, modeling stations, sections, turnouts, and signals as nodes and edges to form an input-capable state graph. The state representation and self-supervised learning module is communicatively connected to the track topology modeling module. It is equipped with a track state representation model. The track state representation model uses a graph neural network to encode the state diagram, extracts key track state features, and trains the track state representation model in combination with a self-supervised learning mechanism. The AlphaEvolve intelligent search and scheduling module is communicatively connected to the state representation and self-supervised learning module. It combines Monte Carlo tree search with policy-value network guidance to search for key track state features, simulate railway signal scheduling, and generate a preliminary scheduling strategy. The evolution strategy fusion module is communicatively connected to the AlphaEvolve intelligent search and scheduling module. It introduces multiple candidate schedulers using a preliminary scheduling strategy, evaluates the strategy performance asynchronously and in parallel, and performs mutation and screening based on reward signals to obtain new samples for scheduling screening. The scheduling rule constraint module is communicatively connected to the AlphaEvolve intelligent search scheduling module and the evolution strategy fusion module, respectively, integrating railway business rules and embedding a scheduling decision mechanism; The output interface module is connected to the AlphaEvolve intelligent search and scheduling module and the evolution strategy fusion module respectively. It obtains the optimal scheduling scheme based on the preliminary scheduling strategy and the new samples selected for scheduling and outputs it.
[0029] The beneficial effects of this invention are: 1. Compared with traditional railway scheduling methods based on rules or single heuristic algorithms, this invention has stronger learning and generalization capabilities and can continuously adapt to different operating scenarios. Self-supervised learning reduces reliance on manual annotation, effectively improving model building efficiency; By integrating evolutionary strategies with enhanced search, a balance between scheduling quality and search efficiency can be achieved, thereby improving the system's practicality. It is suitable for various railway application scenarios such as CTC dispatching platform, ATS automatic control system, and signal simulation platform, and has good engineering deployability.
[0030] 2. Integrating the AlphaEvolve algorithm with evolutionary strategies Based on AlphaEvolve, this approach achieves efficient exploration of complex scheduling spaces through induced policy mutation and population selection. Unlike traditional reinforcement learning, it balances multi-objective optimization with the ability to handle non-differentiable tasks.
[0031] 3. Introduce self-supervised learning to improve state representation. By utilizing unlabeled data from train tracks, state features are learned through self-supervised coding. This eliminates the need for manual labeling, improves generalization ability, and is particularly suitable for high-density railway networks with diverse and complex conditions.
[0032] 4. Multi-module integration—Policy Network + Value Network + Self-supervised GNN By integrating graph neural networks (GNNs) with policy and value networks, scheduling not only considers probabilistic priors but also evaluates the value of current scheduling decisions in real time, thereby achieving more stable and efficient MCTS guidance.
[0033] 5. Applicable to high-density railway networks and complex scenarios The method not only supports high-density train operation, but also has the capability to handle complex scenarios such as multi-node and marshalling yard operations. This represents a significant breakthrough for existing methods in terms of scalability and scenario adaptability.
[0034] 6. Unsupervised state modeling + emphasis on generalization and transfer capabilities A self-supervised mechanism is adopted to enhance the ability to perceive unseen states, reduce reliance on online training, and improve the overall system's adaptability to environmental changes and network topology adjustments.
[0035] 7. Improve the intelligence and response speed of scheduling. Multi-module fusion and evolution strategies accelerate the search, significantly shorten the scheduling response time, and are more suitable for real-time signal adjustment and rapid decision-making scenarios.
[0036] 8. Enhance scalability and system reliability The model is applicable to railway systems of different densities and scales, far exceeding traditional scheduling systems designed for single scenarios, thus reducing deployment and maintenance costs.
[0037] 9. Improve scheduling quality and reduce conflicts and delays. Self-supervised training improves state understanding, and evolution + MCTS deep search decision accuracy can effectively reduce operational conflicts and train delays.
[0038] 10. Save on manual annotation costs and improve data utilization. Self-supervised learning makes full use of existing trajectories and simulation data, eliminating the need for large-scale manual annotation, saving costs and enabling continuous optimization.
[0039] 11. Possesses potential for promotion and cross-domain expansion. The method is highly versatile and adaptable to various transportation networks such as high-speed railways, urban subways, and marshalling yard scheduling. It is expected to be extended to other logistics scheduling fields. Attached Figure Description
[0040] Figure 1 This is a flowchart of the overall process of the method of the present invention. Detailed Implementation
[0041] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention.
[0042] Example 1 Railway signal scheduling optimization method based on AlphaEvolve algorithm, such as Figure 1 As shown, it includes the following steps: Step 1: Orbit State Initialization Obtain the current railway operation diagram, train location, and equipment status; construct the track topology diagram and generate the state diagram structure G(V,E); The initial state node of the initial scheduling is set as the root node of MCTS.
[0043] Step 2: State Coding and Policy Initialization The current state graph is encoded using a GNN to extract key track state features; an initial scheduling probability distribution is generated by a policy network; and the potential scheduling quality of the current state is estimated by a value network.
[0044] Step 3: Monte Carlo Tree Search (MCTS) process Selection phase: Starting from the root node, select child nodes according to the PUCT formula (fusion strategy network output and access count). Expansion phase: If the selected node is not fully expanded, a new child node is added to represent the new scheduling decision; Simulation phase: A complete scheduling simulation is performed starting from this node (simulating trains passing through the path in sequence and determining conflict / success). Feedback phase: The simulation results (such as whether there is a conflict, path cost, delay cost, etc.) are fed back to the nodes on the path to update their estimates.
[0045] Step 4: Evolutionary Strategy-Guided Sample Generation After generating the initial scheduling policy in MCTS, multiple candidate scheduling sequences are introduced as a population; new samples are generated using evolutionary strategies (such as policy perturbation and path mutation). The filtering and updating are based on path costs (such as total train delays and number of operational conflicts).
[0046] Step 5: Update the state model using a self-supervised learning mechanism. The large amount of state transition data generated during the simulation process is used to train the orbital state representation network under self-supervised training. Improve the model's ability to perceive unseen states and its scheduling generalization ability.
[0047] Step 6: Output the optimal scheduling result Based on the combined results of MCTS search results and evolutionary strategy optimization, the scheduling scheme with the fewest conflicts and the least delay is selected. The output includes: train route, passing time table, platform / turnout occupancy sequence, etc.
[0048] The method of the present invention has the following characteristics: Intelligence: Gradually improve the quality of scheduling strategies through learning-based search; Adaptability: Automatically generates scheduling schemes based on real-time status, adapting to different timetables, train types, and emergencies; Scalability: Adaptable to collaborative scheduling scenarios involving multiple stations, multiple lines, and multiple trains.
[0049] This method uses the railway network diagram as the state space basis and the train scheduling sequence as the search target. It comprehensively considers factors such as signal section occupancy constraints, spatiotemporal conflict restrictions, train priority, and platform capacity to optimize operational efficiency under the premise of scheduling safety.
[0050] Example 2 The railway signal scheduling optimization system based on the AlphaEvolve algorithm mainly includes the following key modules: 1. Track Topology Modeling Module The railway signaling system is represented by a graph structure, with stations, sections, turnouts, and signals modeled as nodes and edges to form inputtable state graph data.
[0051] 2. State Representation and Self-Supervised Learning Module By using graph neural networks (GNNs) to encode the orbital graph state and combining it with a self-supervised learning mechanism to train the representation model, deep feature representations of the orbital state can be obtained without a large amount of manually labeled data.
[0052] 3. AlphaEvolve Intelligent Search and Scheduling Module Combining the four steps of traditional MCTS (selection, expansion, simulation, and backpropagation) with the policy-value network-guided strategy, and introducing an evolutionary strategy for breadth exploration and local optimization of the solution space.
[0053] 4. Evolutionary Strategy Fusion Module Multiple candidate schedulers are introduced to evaluate the strategy performance asynchronously and in parallel. Mutation and selection are performed based on reward signals to improve search diversity and escape local optima.
[0054] 5. Scheduling rule constraint module Integrate railway business rules, such as: inter-station timetable constraints, section occupancy restrictions, minimum departure intervals, train priority, timetable adjustment strategies, etc., and embed them into the scheduling decision mechanism.
[0055] 6. Output Interface Module Output scheduling results, including train routes, transit times, and station stopping order, for use by the signaling system or CTC / ATS platform.
[0056] Example 3 This embodiment further elaborates on the above embodiments, explaining the Monte Carlo Tree Search (MCTS) process: I. Selection Stage A) Process Description: Starting from the root node of the search tree (current track state), select the optimal scheduling action path among the child nodes according to the PUCT strategy until a leaf node or a node that has not yet been expanded is reached.
[0057] B) Strategies applied in railway dispatching: Node representation: Each node represents a train scheduling status (e.g., a train has been scheduled, a signal section has been allocated); Action selection considerations: the current set of trains available for dispatch (unoccupied signal sections, no route conflicts); the priority of each train (e.g., passenger trains are higher than freight trains); and the availability of the current signal section. The prior probability π provided by the policy network and the number of visits N are used for balancing.
[0058]
[0059] In the formula, A comprehensive evaluation of the action; For the value of the action; To explore and utilize the equilibrium constant; This represents the initial scheduling probability distribution; This represents the number of times the parent node has been visited. This represents the number of times the action is accessed.
[0060] II. Expansion Phase A) Process Description: If the currently selected node is not fully expanded, add a child node representing a new scheduling action (such as scheduling a train to pass through a signal section).
[0061] B) Rules applied in railway dispatching: For a given scheduling state, the actions that can be extended include: scheduling a train to enter a certain signal section; assigning a train the next running route; and adjusting the departure interval between trains. Business constraints need to be considered: preventing two trains from occupying the same block section at the same time; hard constraints such as speed limits and turnout status. III. Simulation Phase A) Process Description: Starting from the newly expanded child nodes, perform a complete scheduling simulation based on the current state until the end state (such as all trains have been scheduled or the maximum time window has been reached), and evaluate the scheduling quality.
[0062] B) Simulation process applied to railway dispatching: A policy network is used to guide the simulation direction, but a certain degree of randomness is added to maintain diversity. The simulation generates the scheduling trajectory, including: train running path; passing times of each signal; and records of occupied sections. Output simulation evaluation values: Negative factors: number of conflicts, delay time, path deviation; Positive factors: overall traffic efficiency, and good alignment with traffic schedules; Generate a simulated reward value (e.g., -3 indicates 3 conflicts, +10 indicates efficient conflict-free scheduling). IV. Backhaul Phase A) Process Description: The reward values obtained during the simulation phase are uploaded along the path to all traversed nodes to update the average value (Q) and number of visits (N) of each node. The upload formula is as follows:
[0063]
[0064] in, For the first The number of visits to each node; For the first The average value of each node; This is an estimate of the state value; B) Meaning applied to railway dispatching: A higher node access frequency indicates a more optimal scheduling path; A high Q value indicates that the corresponding scheduling action has a good historical performance (few conflicts and high efficiency); Backhaul can enhance high-quality scheduling paths and suppress inefficient or insecure paths.
[0065] Example 4 This embodiment further elaborates on the above embodiments. The execution flow for generating evolution strategy (ES) guided samples in railway scheduling is as follows: I. Initialize the candidate individual population Several high-scoring scheduling sequences {P1, P2, ..., P} were obtained from MCTS. n} Each individual contains: The route of each train (e.g., [S1→S3→S4]); Departure time schedule (e.g., T1 = 08:00, T2 = 08:02); Signal occupancy plan (including turnout allocation, etc.).
[0066] II. Individual variations generate new samples Use the following perturbation operators for individual mutation: Time disturbance: Slightly adjust train departure / arrival times (±30s); Path replacement: Select an equivalent branch from the feasible paths; Sequence rearrangement: Adjusting the dispatching order of two trains; Sectional exchange: exchanging the train arrival sequence at a certain junction; III. Fitness Function Calculation Consider the following multi-objective approach: f 1: Number of signal collisions (the fewer the better); f 2: Total train delay time; f 3: Does it meet the original operation plan (map matching degree)? f 4: Passenger priority weighting (passenger transport is higher than freight transport); Final fitness function:
[0067] in, This is an adjustable coefficient that reflects the scheduling strategy tendency.
[0068] IV. Selection and Feedback of Outstanding Individuals The top k% of scores are selected as winners: Used to update the policy network (supervised learning objective); New actions used to expand nodes on the MCTS tree; Used as the initial individual for the next round of evolution.
[0069] The embodiments of the present invention have been described in detail above, but the present invention is not limited to the described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalents or substitutions are all included within the scope defined by the claims of the present invention.
Claims
1. A railway signal scheduling optimization method based on the AlphaEvolve algorithm, characterized in that, Includes the following steps: S1. Construct the track state diagram and initialize the state diagram; S2. Construct an orbital state representation model, a policy network, and a value network; wherein, the orbital state representation model is constructed based on a graph neural network and is used to encode the state graph and extract key orbital state features; the policy network is used to generate an initial scheduling probability distribution; and the value network is used to estimate the potential scheduling quality of the current state. S3. Construct a Monte Carlo tree and use the selection, expansion, simulation, and backpropagation stages of the Monte Carlo tree search to perform the search. During the search process, for each state node in the tree, extract its key orbital state features using the orbital state representation model, and evaluate the state node through the policy network and value network to guide the search. After the search, generate a preliminary scheduling policy and use the state transition data generated during the simulation process to train the orbital state representation model in a self-supervised manner. S4. Introduce multiple candidate scheduling sequences from the initial scheduling strategy as a population, and use an evolutionary strategy to generate new scheduling samples. Based on path cost, screen the new scheduling samples and update the strategy network. The screened new scheduling samples and the initial scheduling strategy form candidate samples. S5. Select the scheduling scheme with the fewest conflicts and the least delay from the candidate samples as the optimal scheduling scheme and output it.
2. The railway signal scheduling optimization method based on the AlphaEvolve algorithm as described in claim 1, characterized in that, In step S3, the Monte Carlo tree search includes: Selection Phase: Using the initialized track state diagram as the root node, construct a search tree. Starting from the root node of the search tree, select the optimal scheduling action path among the child nodes according to the PUCT formula, and integrate the policy network output and the number of visits. Expansion phase: If the selected child node is not fully expanded, a new child node is added to represent the new scheduling decision; Simulation phase: A complete scheduling simulation is performed starting from this new child node; the scheduling simulation includes simulating trains passing through paths in sequence and determining conflicts / successes; Backhaul phase: The simulation results are backhauled to the nodes on the path to update the node estimates; wherein, the simulation results include whether there is a conflict, path cost, and delay cost.
3. The railway signal scheduling optimization method based on the AlphaEvolve algorithm as described in claim 1, characterized in that, The selection phase of Monte Carlo tree search includes: The selection phase process includes: starting from the root node of the search tree, selecting the optimal scheduling action path among the child nodes according to the PUCT strategy, until a leaf node or a node that has not yet been expanded is reached; The strategies applied to railway scheduling during the selection phase include: Node representation: Each node represents a train scheduling status; Action selection considers: the current set of trains available for dispatch, the priority of each train, and the availability of the current signal section; The initial scheduling probability distribution and the number of visits provided by the policy network are used for balancing, including: In the formula, A comprehensive evaluation of the action; For the value of the action; To explore and utilize the equilibrium constant; This represents the initial scheduling probability distribution; This represents the number of times the parent node has been visited. This represents the number of times the action is accessed.
4. The railway signal scheduling optimization method based on the AlphaEvolve algorithm as described in claim 1, characterized in that, The extended phases of Monte Carlo tree search include: The expansion phase process description includes: if the currently selected node is not fully expanded, add a child node representing the new scheduling action to it; The rules applied to railway scheduling during the extended phase include: For a scheduling state, the actions that can be extended include: arranging a train to enter a certain signal section, assigning a train the next running path, and adjusting the departure interval between trains; Business constraints need to be considered: preventing two trains from occupying the same block section at the same time, and hard constraints on speed limits and turnout status.
5. The railway signal scheduling optimization method based on the AlphaEvolve algorithm as described in claim 1, characterized in that, The simulation phase of Monte Carlo tree search includes: The simulation phase process description includes: starting from the newly expanded child nodes, performing a complete scheduling simulation based on the current state until the end state, and using the value network to evaluate the scheduling quality; The simulation process applied to railway dispatching includes: Use a policy network to guide the simulation direction, and add a certain degree of randomness to maintain diversity; The simulation generates the scheduling trajectory, including: train running path, passing time of each signal, and occupied section record; Output simulated evaluation values: Negative factors: number of conflicts, delay time, and route deviation; Positive factors: overall traffic efficiency, and good alignment with traffic schedules; A simulated return value is generated by combining all the data.
6. The railway signal scheduling optimization method based on the AlphaEvolve algorithm as described in claim 1, characterized in that, The return phase of the Monte Carlo tree search includes: The backhaul phase process includes: uploading the reward values obtained in the simulation phase along the path to all traversed nodes to update the average value Q and access count N of each node. The backhaul formula is: in, For the first The number of visits to each node; For the first The average value of each node; This is an estimate of the state value; The meaning of the feedback phase in railway dispatching includes: A higher node access frequency indicates a more optimal scheduling path; A high Q value indicates that the corresponding scheduling action has a good historical performance; Backhauls enhance high-quality scheduling paths and suppress inefficient or insecure paths.
7. The railway signal scheduling optimization method based on the AlphaEvolve algorithm as described in claim 1, characterized in that, Step S4 includes: S41. Initialize the candidate individual population: Obtain several high-scoring scheduling sequences from Monte Carlo tree search as the population. Each individual in the population contains: the running path of each train, the departure time plan, and the signal occupancy plan. S42. Individual mutation to generate new samples: Use perturbation operators to perform individual mutation to generate new scheduled samples; S43. Fitness function calculation: Using the fitness function, calculate the score of scheduling new samples under multi-objective conditions; S44. Selection and Feedback of Winning Individuals: Select the top k% of the new scheduling samples as winners, and use the winners to update the policy network, to expand the new actions of nodes on the Monte Carlo tree search tree, and to serve as the initial individuals for the next round of evolution.
8. The railway signal scheduling optimization method based on the AlphaEvolve algorithm as described in claim 7, characterized in that, In step S42, the perturbation operator includes: Time disturbance: Minor adjustments to train departure / arrival times; Path replacement: Select an equivalent branch from the feasible paths; Sequence rearrangement: Adjusting the dispatching order of two trains; Sectional exchange: exchanging the train arrival order at a certain junction.
9. The railway signal scheduling optimization method based on the AlphaEvolve algorithm as described in claim 7, characterized in that, In step S43, the fitness function is: in, For new samples The score, This represents the number of signal collisions. Total train delay time; To determine whether the original operating schedule is met; Weight passengers according to their priority. , , , This is an adjustable coefficient that reflects the scheduling strategy tendency.
10. A railway signal scheduling optimization system based on the AlphaEvolve algorithm, which is based on the railway signal scheduling optimization method of any one of claims 1-9, characterized in that, include: The track topology modeling module uses a graph structure to represent the railway signaling system, modeling stations, sections, turnouts, and signals as nodes and edges to form an input-capable state graph. The state representation and self-supervised learning module is communicatively connected to the track topology modeling module. It is equipped with a track state representation model. The track state representation model uses a graph neural network to encode the state diagram, extracts key track state features, and trains the track state representation model in combination with a self-supervised learning mechanism. The AlphaEvolve intelligent search and scheduling module is communicatively connected to the state representation and self-supervised learning module. It combines Monte Carlo tree search with policy-value network guidance to search for key track state features, simulate railway signal scheduling, and generate a preliminary scheduling strategy. The evolution strategy fusion module is communicatively connected to the AlphaEvolve intelligent search and scheduling module. It introduces multiple candidate schedulers using a preliminary scheduling strategy, evaluates the strategy performance asynchronously and in parallel, and performs mutation and screening based on reward signals to obtain new samples for scheduling screening. The scheduling rule constraint module is communicatively connected to the AlphaEvolve intelligent search scheduling module and the evolution strategy fusion module, respectively, integrating railway business rules and embedding a scheduling decision mechanism; The output interface module is connected to the AlphaEvolve intelligent search and scheduling module and the evolution strategy fusion module respectively. It obtains the optimal scheduling scheme based on the preliminary scheduling strategy and the new samples selected for scheduling and outputs it.