Enhancing Multiagent Genetic Network Programming Performance Using Search Space Reduction

Ali Kohan,Mohamad Roshanzamir,Roohallah Alizadehsani

2024-12-15

Abstract:Genetic Network Programming (GNP) is an evolutionary algorithm that extends Genetic Programming (GP). It is typically used in agent control problems. In contrast to GP, which employs a tree structure, GNP utilizes a directed graph structure. During the evolutionary process, the connections between nodes change to discover the optimal strategy. Due to the large number of node connections, GNP has a large search space, making it challenging to identify an appropriate graph structure. One way to reduce this search space is by utilizing simplified operators that restrict the changeable node connections to those participating in the fitness function. However, this method has not been applied to GNP structures that use separate graphs for each agent, such as situation-based GNP (SBGNP). This paper proposes a method to apply simplified operators to SBGNP. To evaluate the performance of this method, we tested it on the Tileworld benchmark, where the algorithm demonstrated improvements in average fitness.

Multiagent Systems

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is **the optimization difficulties in Genetic Network Programming (GNP) in multi - agent systems due to the overly large search space**. Specifically, GNP is an evolutionary algorithm that extends Genetic Programming (GP), and it uses a directed - graph structure instead of a tree structure to represent programs. This structure enables GNP to handle complex problems more effectively, but at the same time, it also leads to an extremely large search space, especially in the case where each agent uses an independent graph structure, such as Situation - Based GNP (SBGNP). This makes it very difficult to find the optimal solution.

To address this challenge, the author proposes a new method to reduce the search space through **simplified operators**. These simplified operators limit the changes in node connections, only allowing those nodes involved in the fitness function calculation to change. This method can significantly reduce the search space, thereby improving the convergence speed and efficiency of the algorithm.

### Summary of Main Problems:
1. **Excessively Large Search Space**: Due to its directed - graph structure, GNP has a very large search space, especially when using an independent graph structure in a multi - agent environment.
2. **Optimization Difficulties**: The large search space makes it very difficult to find the optimal solution within a limited number of iterations.
3. **Insufficiency of Existing Methods**: Although the existing simplified operator methods can reduce the search space, they have not yet been applied to GNP variants (such as SBGNP) in a multi - agent environment.

### Solutions:
- **Application of Simplified Operators**: By limiting the changes in node connections and only allowing the branches accessed at runtime to participate in evolutionary operations (crossover and mutation), the search space is effectively reduced.
- **Application in a Multi - Agent Environment**: Apply the simplified operators to Situation - Based GNP (SBGNP) to improve its performance in a multi - agent environment.

### Experimental Verification:
The author evaluated the proposed algorithm in the Tileworld benchmark test environment and compared it with standard GNP, GNP_simplified, and SBGNP. The experimental results show that the new method performs well in terms of average fitness and the number of successful solutions, especially having more obvious advantages in more complex environments.

Through these improvements, this research provides an effective method for improving the performance of multi - agent GNP, reducing the search space and improving the optimization efficiency.