Planning and execution of actions for agents in dynamic environments

点击翻译按钮获取中文摘要

In artificial intelligence, an agent is a computer program capable of acquiring information about its environment in order to perform purposeful tasks autonomously. It could be a piece of software controlling an autonomous robot, using sensors to perceive the environment, making inferences to interpret the percepts, as well as planning and executing actions to achieve goals. It could also be the 'Artificial Intelligence' or 'AI' controlling units in a Real-Time Strategy (RTS) game, observing the moves made by a player to decide how to react in order to offer an entertaining challenge. Endowing such an agent with the capability to autonomously plan and execute actions is is still a tremendous challenge. Such a planning and plan execution capability should be coordinated with the agent's capabilities to perceive and act, taking into account its state, that of the environment, and the goal to be achieved. Today most efficient planning algorithms use plan representations with limited expressiveness, as well as heuristics and search-control methods geared towards deterministic environments. When these algorithms are used in dynamic environments such as RTS games or mobile robots, the resulting planning capability is very limited. Furthemore, current methods for coordinating planned and reactive actions for such agents have a limited effectiveness. To address these problems, this project will develop new planning algorithms, as well as new methods for coordinating the execution of planned and reactive behaviours.