| AI planning forms a sequence of actions through reasoning and searching to achieve specific goals.The core of AI planning research lies in improving the efficiency and quality of problem-solving,where heuristic search plays an extremely important role.The development of a heuristic function and search policy is a key component of heuristic search,where a good heuristic function and efficient search algorithm can help find the optimal solution in large-scale search spaces quickly.However,the current research on heuristic search faces some challenges.On the one hand,existing heuristic functions do not provide sufficient heuristic information,and their estimated values may deviate from the true cost,making it difficult to accurately guide the search.On the other hand,traditional heuristic search algorithms have high space costs when dealing with large-scale domain problems and are overly dependent on heuristic functions,leading to inefficient problem-solving.Neural network has powerful learning capabilities that can be applied in various aspects of planning to improve problem-solving efficiency.The focus of this study is on how to use neural networks to learn optimized heuristic function and how to utilize their generalization ability to formulate a generalized search policy for a domain.This thesis focuses on the characteristics of state nodes in planning problems and investigates the sub-goal partition,heuristic function optimization,action importance learning,and search policy design.In addition,this thesis provides a solution to the emergency material scheduling problem in the application domain.The research in this thesis includes the following three aspects:Firstly,for the optimization of heuristic function,this thesis proposes a sub-goal partition algorithm to divide the goal atoms for optimizing heuristic function.The partition information is then applied to feature encoding.By mapping planning problems to relaxed problems and constructing hypergraph data,a graph neural network is utilized to learn a heuristic function.Additionally,a search algorithm suitable for dual heuristic functions is proposed to evaluate its stability.Experimental results demonstrate that in most domains,the learned heuristic function has a success rate of over 80%,which reduces the search space and time compared to traditional heuristic function,resulting in higher problem-solving capability and stability.Secondly,for generalized search policy,this thesis uses canonical abstraction to encode features of states and constructs a fully connected network model to learn the importance of action schema,as well as a graph neural network model to learn object importance.The action is evaluated based on the importance of action schema and object,and the A~*search algorithm is improved based on action importance to form generalized search policy.Compared with A star search,experimental results show that the proposed method overcomes the problem-solving of low efficiency caused by inaccurate heuristic function estimation.For some problems in specific domains,the number of expanded nodes and the solving time significantly decrease.Finally,for emergency material scheduling problems under unexpected circumstances,this thesis analyzes social demands and proposes an AI planning model.By describing the traffic conditions and material demand locations under unexpected circumstances,combining heuristic search algorithms and numerical planning,the planning system can effectively guide multiple parties to participate in emergency material dispatching,satisfying emergency material demands as soon as possible. |
