Conclusions

The goal of this project was to demonstrate the applicability of the DITOPS scheduler to the problem of reactive re-planning in the medical evacuation domain. We believe that this goal has clearly been achieved. In a relatively short period of time, a quite sophisticated planner was designed and built. The planner incorporates important domain constraints and assumptions (e.g., multi-leg patient itineraries, finite transport capacity, finite in-transit (ASF) and destination (MSF) medical treatment capacity), is capable of revising medical evacuation plans in response to a substantial set of disruptive events, and appears directly scalable to full-scale evacutation problems and scenarios. The planner provides strong evidence of the utility of the modeling and scheduling componentry provided within DITOPS, and a compelling example of the efficacy of approaching application development as a differential process of component configuration and customization. This approach, in our view, is pre-requisite to cost-effective development of complex planning and scheduling applications.

Evaluating the plans created by the DITOPS medical evacuation planner is not an easy task. The system generates and maintains feasible plans (modulo the system's current constraint models). However, due to the unavailability of both realistic input data and comparative results from other medical evacuation planning systems, it has not yet been possible to systematically evaluate the system's problem solving performance, or to determine how well the planner would do solving actual medical evacuation problems (the problems it has solved are only our understanding of what the real world is like). The next logical step in this research would be to conduct some sort of analysis of current system capabilities. One starting point could be comparison to current TRACES planning and replanning modules on benchmark problems such as the "Lilliput" scenario (this problem has been solved by the current prototype). Evaluation and feedback from the medical evacuation planners would also seem essential at this point. To facilitate such user involvement and evaluation, the current prototype provides a graphical interactive framework for flexibly experimenting with various disruptive events and examining the results of reactive responses.

One obvious direction for further development of the current medical evacuation planner would be to expand the range of constraints that are accommodated and enforced within the system's domain model. Our initial effort has concentrated attention on an identified subset of "important" domain constraints. To be practically applicable, a re-planning capability would necessarily need to take into account additional aspects of the domain. Given our current understanding of the medical evacuation re-planning problem, we can identify several domain details and constraints that are ignored in the current prototype:

• Difference between cargo and medical missions. Distinguishing between these two would not be difficult to add.
• Range of aircraft missions. DITOPS models aircraft range but the range values are not used in this implementation; taking range into account when adding new missions/legs would be straightforward.
• Maximum flying altitude and other specific request for patients (like medical equipment and other requirements).
• Litter seats vs. ambulatory seats: all seats are considered to be the same. To implement this, DITOPS provides a mechanism for partitioning capacity of resources; in this case two partitions would be used.
• Different aircraft configurations have not been modeled.
• Patients' airfield preferences are not enforced (personal preferences like planning for the destination to be close to his/her home town etc.).
• Ground transportation is not provided.
• Missions are not modified to pick up urgent patients.
• Crew requirements are not modeled - the maximum flight time should be enforced by somebody else. We do not enforce any constraint on given missions besides spatial and temporal constraints.
• Compatibility aspects relating airfields and aircraft are not considered. DITOPS has a mechanism called compatibility constraints which could be used to model this.
• Secondary resources (like fuel and food) are not modeled.

Another direction of future development would be the introduction of more sophisticated planning and scheduling strategies. The current medical evacuation planner makes little use of underlying architectural capabilities for conditionalizing and integrating the use of different solution repair methods, but rather associates a particular repair method with each type of disruption or conflict. Consider, for example, the introduction of a mission delay, which in turn causes conflicts with respect to connecting flights in several patient itineraries. Currently, the planner resolves such conflicts by reassigning each affected patient to a new (feasible) connecting flight. However, there are tradeoffs that could be considered here. If a conflict involving a connecting flight is quite small (e.g., the flight is scheduled to leave just minutes too soon), then a much more reasonable reaction might be to simply delay the connecting flight. In general, there are opportunities for improved reactive planning behavior through increased analysis of the constraints involved in recognized conflicts and more selective use of constituent repair methods. Likewise, there are also straightforward extensions to the current set of methods which could provide a basis for more effective localized reaction in some circumstances. A version of the patient scheduler which allows selective "bumping" of currently scheduled patients (e.g., to allow the rescheduling of an urgent patient to displace a non-urgent patient if necessary) would be one obvious extension that is easily implemented.