Robots for rescue operations after a disaster are an interesting and challenging research problem that has the potential to save lives and reduce economic losses after a disaster. We developed TOPTWR, an extension of the popular TOPTW model, to model the issues in task allocation for teams of rescue robots. Our hybrid algorithm is based on a team of heterogeneous humanoid robots trying to optimize five objectives (task rewards, task completion time, total energy, maximum energy consumption for a single robot, and missed deadline penalties). A common approach to solve these kinds of problems are multi-objective evolutionary algorithms (MOEAs), but their major disadvantage is that they cannot deal with dynamic environments easily. This paper presents an efficient solution for TOPTWR by combining MOEAs with learning algorithms. A novel Extended Multi-Start Simulated Annealing Iterated Local Search (EMSAILS) operator using a modern state-of-the-art NSGA-III algorithm is proposed. In addition, we applied Q-Learning to learn the likely changes in the environment and how to react to them. This algorithm, HMO-TOPTWR-NSGA-III (HMO-N-L), uses an artificial neural network (ANN) as a function approximator to make the huge state and action spaces tractable. This paper includes a thorough empirical evaluation demonstrating the effectiveness of the multi-objective algorithm in both static and dynamic environments. The evaluation shows that the proposed algorithm reduces the error by up to 42% against three state-of-the-art approaches to TOPTW (HMO-N, MSA, and IPI).

灾难后用于救援行动的机器人是一个有趣且具有挑战性的研究问题，有可能挽救生命并减少灾难后的经济损失。我们开发了TOPTWR，这是流行的TOPTW模型的扩展，为救援机器人团队的任务分配中的问题建模。我们的混合算法基于一组异类人形机器人，它们试图优化五个目标（任务奖励，任务完成时间，总能量，单个机器人的最大能量消耗以及错过最后期限的惩罚）。解决这类问题的常用方法是多目标进化算法（MOEA），但是它们的主要缺点是它们无法轻松应对动态环境。本文通过将MOEA与学习算法相结合，提出了一种有效的TOPTWR解决方案。提出了一种使用现代NSGA-III算法的新型扩展多起点模拟退火迭代本地搜索（EMSAILS）算子。此外，我们应用了Q-Learning，以了解环境中可能发生的变化以及如何应对。该算法HMO-TOPTWR-NSGA-III（HMO-NL）使用人工神经网络（ANN）作为函数逼近器，以使庞大的状态空间和动作空间易于处理。本文包括全面的实证评估，证明了多目标算法在静态和动态环境中的有效性。评估显示，与三种最先进的TOPTW方法（HMO-N，MSA和IPI）相比，所提出的算法可将错误减少多达42％。