A penalized batch-Bayesian approach to informative path planning for decentralized swarm robotic search
A penalized batch-Bayesian approach to informative path planning for decentralized swarm robotic search
复制标题
DOI：
10.1007/s10514-022-10047-8
复制
发表时间：
2022-06
影响因子：
3.5
通讯作者：
P. Ghassemi;Mark Balazon;Souma Chowdhury
中科院分区：
文献类型：
--
作者：
P. Ghassemi;Mark Balazon;Souma Chowdhury
用于搜索和目标定位的群体机器人方法，其中目标源发出空间变化的信号，保证了无与伦比的时间效率和鲁棒性。对于大多数现有的群体搜索方法，同时保持搜索效率和数学洞察力以及可扩展性和计算易处理性仍然具有挑战性。我们最近开发的分散方法 Bayes-Swarm-O 是一种基于批量贝叶斯优化的基于模型的方法，已被证明在搜索效率方面优于最先进的群体启发式方法。然而，这种原始的贝叶斯群体方法没有考虑机器人决策之间的相互作用（又名批量样本），并且被发现对探索和探索之间的规定平衡敏感。本文缓解了这些限制，通过分别使用新的边缘化惩罚方法来体现批量采样和任务期间探索/利用平衡的动态适应，显着提高了搜索效率和收敛性。此外，本文还介绍了通过基于 Pybullet 的新型分布式群体搜索模拟器执行的一组系统实验，分析了群体大小增加、部分同行观察和优化器选择对这种更新算法（现在称为 Bayes-Swarm-P）的影响。在模拟多模态信号分布和滑雪者/雪崩搜救问题上，与三种标准群体搜索方法（即 Glowworm 搜索、Levy 行走和穷举搜索）相比，先进的 Bayes-Swarm-P 方法在搜索效率和鲁棒性方面也明显优越。
Swarm-robotic approaches to search and target localization, where target sources emit a spatially varying signal, promise unparalleled time efficiency and robustness. With most existing swarm search methods, it remains challenging to simultaneously preserve search efficiency and mathematical insight along with scalability and computational tractability. Our recently developed decentralized method,Bayes-Swarm-O, a model-based approach founded on batch Bayesian Optimization, has been shown to outperform state-of-the-art swarm heuristics in terms of search efficiency. However, this originalBayes-Swarm-Omethod did not account for the interactions between robots’ decisions (aka samples in a batch) and was found to be sensitive to the prescribed balance between exploration and exploration. These limitations are alleviated in this paper, leading to significantly improved search efficiency and convergence, by respectively using a new marginalization penalization approach to embodied batch sampling and a dynamic adaptation of the exploration/exploitation balance during mission. In addition, this paper presents a systematic set of experiments executed through a new Pybullet-based distributed swarm search simulator, that analyzes the impact of increasing swarm size, partial peer observation, and choice of optimizer, on this updated algorithm, now calledBayes-Swarm-P. The advancedBayes-Swarm-Pmethod is also found to be clearly superior in terms of search efficiency and robustness when compared to three standard swarm search methods (namely Glowworm search, Levy walk, and exhaustive search) over simulated multimodal signal distributions and a skier/avalanche search and rescue problem.