You are here

Learning Performance Models and Tactical Knowledge for Continuous Mission Planning

Description:

TECHNOLOGY AREA(S): Info Systems, Ground Sea, Battlespace 

OBJECTIVE: Develop machine learning approaches for automatically acquiring and continually updating asset performance models and tactical planning knowledge to improve the decision support by automated mission planning systems in highly dynamic environments, and to enable their maintainability. 

DESCRIPTION: Automated planning tools are commonly used as decision aids for mission planners. Successful mission planning requires accurate and complete models of the performance capabilities of the assets, the environment including the behaviors of other agents in the environment, and mission goals and sub-goals. Current practice for planning in situations that change is to hand-code the changes in the models of capability, environment, and goals and then re-plan. This approach is slow and becomes infeasible in highly dynamic situations, particularly in tactical mission planning where the tempo of new information requires rapid changes in the models that may become inconsistent and obsolete faster than our ability to hand-code new models. This can severely degrade the quality and effectiveness of automated planning aids to a degree that they may not be used. This problem is further exacerbated by the introduction of unmanned assets if there are frequent changes to their sensing and autonomous capabilities. The Navy needs to develop methods that can rapidly and continuously plan as new information necessitates updating the models. Machine learning is a promising approach for learning to continually update performance and the environment models for use in automated planning. The Navy wants to develop learning methods that can leverage and exploit mission performance data and user feedback including after action reports, as well as planning decisions and critiques of system performance. Recent advances in machine learning are applicable to these automated mission planning aids, and could allow them to automatically improve their performance with respect to new asset models, and incorporate new protocols appropriate to encountered situations. For instance, the ability to learn and update asset performance models has been demonstrated in certain domains [1]; which could prove useful in learning predictive asset performance models. Likewise, task model learning has been demonstrated with hierarchical task network learning [2][4] and explanation-based learning [3]. Applicable approaches for learning asset performance models and tactical knowledge for use in complex multi-domain, multi-asset mission planning problems that are scalable and robust are desired. Work produced in Phase II may become classified. Note: The prospective contractor(s) must be U.S. Owned and Operated with no Foreign Influence as defined by DOD 5220.22-M, National Industrial Security Program Operating Manual, unless acceptable mitigating procedures can and have been implemented and approved by the Defense Security Service (DSS). The selected contractor and/or subcontractor must be able to acquire and maintain a secret level facility and Personnel Security Clearances, in order to perform on advanced phases of this contract as set forth by DSS and ONR in order to gain access to classified information pertaining to the national defense of the United States and its allies; this will be an inherent requirement. The selected company will be required to safeguard classified material IAW DoD 5220.22-M during the advance phases of this contract. 

PHASE I: To develop and evaluate the feasibility of the approach, proposers should specify a realistic application scenario and assets and their performance characteristics. For example, a scenario of interest is continuous planning for maritime mine detection and neutralization using unmanned vehicles. Conduct a study of machine learning approaches that could be used to acquire and update asset performance models and planning knowledge for automated planning systems. Assess the feasibility of selected approaches for incrementally and continually learning performance and planning knowledge in the context of multi-domain, multi-asset missions. Identify software interface and requirements for integrating learning algorithms with automated planning aids. Phase I should include plans for a prototype to be developed during Phase II. 

PHASE II: Implement machine learning algorithms identified in Phase I into a software prototype. Evaluate the effectiveness of learning over multiple simulated scenarios and systems. Evaluate and demonstrate the effectiveness using measures such as improvement in coverage, increased acceptance of planning recommendations and subsequent increase in mission measures of effectiveness and performance. Work in this phase may be done at the unclassified level; however, the ability to handle restricted databases would add flexibility. It is probable that the work under this effort will be classified under Phase II (see Description section for details). 

PHASE III: Mature and extend the learning algorithms to operate effectively, be robust, and fault-tolerant to a range of government-provided data and constraints, in planning systems under their operating conditions. Coordinate with the program office to fully test and integrate into a potential program of record. Private sector commercial potential and dual-use applications include survey and first responder operations. 

REFERENCES: 

1: Ozisikyilmaz, Memik, Choudhary (2008). Machine Learning Models to Predict Performance of Computer System Design Alternatives. In Proceedings of 37th International Conference on Parallel Processing. http://cucis.ece.northwestern.edu/projects/DMS/publications/OziMem08B.pdf

2:  Garland, Lesh, (2003). Learning Hierarchical Task Models by Demonstration. Technical Report TR2003-01, Mitsubishi Electric Research Laboratories. http://www.cs.brandeis.edu/~aeg/papers/garland.tr2002-04.pdf

3:  Mohan, Laird (2014). Learning Goal-Oriented Hierarchical Tasks from Situated Interactive Instruction. Proceedings of the 27th AAAI Conference on Artificial Intelligence (AAAI). http://web.eecs.umich.edu/~soar/sitemaker/docs/pubs/mohan_AAAI_2014.pdf

4:  Zhuo, Munoz-Avila, Yang (2014). Learning Hierarchical Task Network Domains from Partially Observed Plan Traces. Artificial Intelligence Journal. http://www.cse.lehigh.edu/%7Emunoz/Publications/AIJ14.pdf

KEYWORDS: Continuous Planning; Tactical Mission Planning; Automated Planners; Dynamic Environments; Machine Learning; Learning Performance Capabilities 

CONTACT(S): 

Behzad Kamgar-Parsi 

(703) 696-5754 

behzad.kamgarparsi@navy.mil 

Jason Stack 

(703) 696-2485 

US Flag An Official Website of the United States Government