Welcome to CMU's Multi-Agent Planning, Learning, and Execution (MAPLE) Project. Our research goal is to develop new, effective algorithms for distributed multi-agent control in highly dynamic, physical environments. Such environments pose a variety of challenges:

• Limits in perception and inter-agent communication make it impossible to accurately determine the state of the environment; hence, decisions must be made under uncertainty. Gathering accurate information is often a crucial subgoal of achieving a specific task. Information collected by different agents may be highly inconsistent and incomplete.

• Decisions have to be made in real time. The environment never waits for the termination of a complex planner, or for the input of a human decision maker. Unanticipated events, often arising through changes in the environment or failure of specific agents, may require massive replanning during plan execution.

• Agents must be tied together by a flexible communication and control brokering architecture (the "Grid"). The Grid must fully support distributed decision making, distributed information gathering, and dynamic role allocation to individual physical entities. The Grid must also cope with variable number of agents, and fully support the integration of information and control at all levels of decision making.

CMU's MAPLE project investigates these issues in the context of a specific disaster scenario: Hurricane Mitch. In late 1998, Hurricane Mitch devastated large parts of central America, killing tens of thousand of people and leaving millions homeless. It is our hypothesis that a more effective coordination strategy for forces on the ground and in the air could have saved many lives.

To shed light on this conjecture, we have developed a graphical simulator of Hurricane Mitch and Honduras, based on accurate geographical and climate data collected in the disaster. The MAPLE project has also developed a new family of real-time learning and decision making techniques that use efficient, sample-based algorithms for real-time decision making. Uncertain information is represented by probability distribution, using numerical factors to trade-off different hypotheses based on evidence. Fast, sample-based algorithms are used to generate decisions in an any-time fashion, meaning that answers are available at any time, but the more decision time is available, the better the performance. Currently, these algorithms are adapted to the distributed nature of the Grid, and evaluated using the MAPLE simulator of Hurricane Mitch.

The MAPLE project is carried out in close collaboration with DARPA's MIATA TIE, funded under DARPA's CoABS program (short for: Control of Agent-based Systems). Our joint goal is to demonstrate that new, agent-based technology will lead to a tenfold increase in effectiveness of mixed-initiative planning and execution, at all levels of command and control. The MAPLE simulator and system provide an open API (currently still under construction), enabling others to interact directly with a team of simulated, physical agents in Honduras.