摘要
随着人们对智能体需求的提高,智能体的活动不再局限于简单环境与单一任务,面向复杂的应用场景,智能体需要具备自主完成决策与执行的能力.本文研究了面向线性时序逻辑描述下的复杂任务智能体的不确定行为规划问题,同时考虑任务成功率与任务执行成本,这里不确定性因素包括智能体行为与环境属性,任务描述由软、硬约束两部分表达.文中应用形式化方法中模型检测的思想确定智能体行为选择策略,其中应用马尔可夫决策过程构建智能体个体与集群模型,应用双层自动机模型构建任务模型,设计智能体-任务网络模型表征约束条件并通过一耦合线性规划完成策略求解,并通过数值仿真测试对上述方法进行了验证.结果表明含软、硬约束的复杂任务约束可被满足,求解所得最优策略使智能体按约束强度完成任务,且可通过调节惩罚因子控制网络模型的松弛程度调整生成的控制策略.
With the increase of people’s demand for agents, the activities of agents are no longer limited to simple environment and task format.Facing complex application scenarios, agents need to be able to make decisions and execute them autonomously. This paper studies the probabilistic action planning considering complex task constraints described by linear temporal logic. At the same time, the success rate and cost of task are both considered. The uncertain factors include agent behavior and environment attributes, and the task description is expressed by soft and hard constraints. The strategy of agent is generated here applying model checking in formal method. Single-and multi-agent model is established using Markov decision process, while task model is established using doublelayer automata. Then, agent-task network model is designed to describe the constraints and the control strategy is solved through a coupled linear programming. The method above is verified through numerical simulation. The results show that the complex task constraints in the form of soft and hard constraints can be satisfied. The optimal strategy enable the agent to complete the task according to the constraint strength, and the control strategy can be adjusted by controlling the relaxation degree of control network model relevant to the penalty factor.
作者
陈仲瑶
方浩
CHEN ZhongYao;FANG Hao(School of Automation,Beijing Institute of Technology,Beijing 100081,China)
出处
《中国科学:技术科学》
EI
CSCD
北大核心
2020年第5期516-525,共10页
Scientia Sinica(Technologica)
基金
深圳机器人基础研究中心项目国家自然科学基金(编号:U1913602)
重大国际(地区)合作研究项目、国家自然科学基金(批准号:61720106011,61873033,61903035)
鹏城实验室和智能机器人与系统高精尖创新中心资助。
关键词
线性时序逻辑
不确定行为规划
双层自动机
软、硬任务约束
linear temporal logic
probabilistic action planning
double-layer automata
soft/hard constraint