چكيده به لاتين
With the expansion of robotic systems, intelligent decision-making and resource allocation management have become necessary and essential. This issue is of greater significance in multi-agent systems due to the increased number of agents. One of the fundamental challenges in this area is the real-time solution of constrained multi-objective optimization problems. In this context, two teams are considered: the friendly team and the adversarial team. Agents in the friendly and adversarial teams are referred to as agents and targets, respectively. The constraints of the problem include operational range, asset network, agents' takeoff times, maximum speeds of each agent, and the number of agents considered. The operational range is the area of activity of the friendly team, while the asset network highlights important areas within this operational range. The objective of this thesis is the optimal allocation of agents from the friendly team to the important targets of the adversarial team. To solve the problem, given the limited number of agents for allocation, we first evaluate the prospective future behavior of the targets. This assessment is based on the proximity to both the operational range and the asset network. The future movement behavior of the targets is determined using a combined approach of reinforcement learning and collective learning, and the targets are prioritized based on their importance. After prioritizing and selecting the most important targets, we allocate agents to these targets by solving the constrained optimization problem while minimizing time. The time required to solve the optimization problem is significant due to the large number of variables for agents and targets, and thus it cannot be effectively used in real-time. To address the problem in real-time, we use collective learning methods to approximate and solve the optimization problem. By treating agents and targets as multi-rotor agents and randomly assigning operational ranges and asset networks, we generate artificial datasets. By combining artificial data with real data, we train the “decision-making system.” To enhance the model's accuracy, we determine the model's hyperparameters using a genetic algorithm. Finally, to solve the problem in real-time, we first identify important targets using a prioritization function. Then, we use the decision-making system to allocate agents to these targets. To evaluate the decision-making system and the prioritization function, we assess the decision-making system using simulation evaluation data and practical test data that were not involved in the training process.
