MODIFICATION
A -- Integrated Command & Control
- Notice Date
- 8/16/2012
- Notice Type
- Modification/Amendment
- NAICS
- 541712
— Research and Development in the Physical, Engineering, and Life Sciences (except Biotechnology)
- Contracting Office
- Department of the Air Force, Air Force Materiel Command, AFRL - Rome Research Site, AFRL/Information Directorate, 26 Electronic Parkway, Rome, New York, 13441-4514
- ZIP Code
- 13441-4514
- Solicitation Number
- BAA-10-01-RIKA
- Point of Contact
- Lynn G. White, Phone: (315) 330-4996
- E-Mail Address
-
Lynn.White@rl.af.mil
(Lynn.White@rl.af.mil)
- Small Business Set-Aside
- N/A
- Description
- Machine Intelligence for Mission Focused Autonomy (MIMFA) Integrated C2 BAA Amendment Autonomous systems have the potential to significantly improve the agility and effectiveness of our nation's military while reducing manpower and cost requirements. However, the robustness of current autonomous and automated control technologies is inadequate given the dynamics and complexities of practical real-world problems. The Machine Intelligence for Mission-Focused Autonomy (MIMFA) program will research technologies that enable distributed autonomous assets to make intelligent coordinated adaptations to plans/policies to optimize mission performance in dynamic and realistically complex domains. Emphasis is placed upon autonomous techniques which enable faster, more efficient understanding for complex, multi-modal military data. The ability to autonomously seek out, understand, and present information to human decision makers will contribute greatly to future mission success. General technical areas for this MIMFA program are Knowledge Discovery and Data Mining (KDD), Machine Learning (ML), and Multi-Agent Systems (MAS). KDD is the capability to draw conclusions, find specific information, and fine-tune understanding by bringing together large, heterogeneous data sources. ML is the capability for software to adapt its behavior through training or experience to better solve problems. MAS are systems of intelligent software agents interacting to solve highly complex problems, collaborate amongst themselves, and observe and present information on behalf of the human. By combining these three technical areas, we believe there are powerful synergies to be found in helping Command Control Intelligence Surveillance Reconnaissance (C2ISR) operators better understand complex problems and take decisive action. The application context of interest is the Data to Decisions domain and its associated processes. Within this domain the MIMFA program envisions an information spectrum containing various states of information within which we foresee the autonomous multi agent systems working. The first is a dynamic information state which includes self-organizing information collectors. Our goal is to enable the capability for a group of heterogeneous agents to seek mission-relevant information. To accomplish this we wish to advance the science and technology related to the tasking, coordination, self-organization, and negotiation of distributed agents in and amongst the complex domains of air, space, and cyber. Interesting, but not mandated, approaches include extending game theory to reason across the dynamics of individual and group utility to drive agent rationality, as well as both distributed constraint satisfaction and auctioning as mechanisms for self-organization without the need for a central control. The second information state is a static state in which information is at rest in any number of large data sets. The resulting capability required in this state is for a group of heterogeneous agents that can examine data from different perspectives, compare multiple hypotheses, classify important states and features, identify gaps in understanding, and ultimately provide a robust and rational analysis for human decision-makers. Performing this analysis in a timely manner relies upon the agents' ability to dynamically self-adapt their capabilities to new data, as well as understanding and building upon the trust and confidence of human operators. A successful group of agents will operate to answer the commander's need for actionable decision options which take into account multiple perspectives on the data and demonstrate a robust understanding of the problem, potential solutions, risks, and information provenance. Relevant technical areas, but NOT of interest to the MIMFA program at this time include: - Trust - Trust in the output of the autonomous agents should be considered during technology development. While the issue of trust in autonomous agent systems is a major concern, it will be addressed specifically in a subsequent solicitation. - Group/Transfer Learning - Group and transfer learning between heterogeneous multi agent systems is of high importance to the MIMFA program, but will be the addressed specifically in a subsequent solicitation. - Human/Robot Interfaces (HRI) - The MIMFA program plans to leverage, to the maximum extent, previous and existing investments in HRI and agent visualization research. Agent feedback to the human operator is an important capability. Offerors looking to propose HRI or visualization as part of their proposal should consider utilizing existing HRI and visualization capabilities. - Robotic Control - The MIMFA program is not interested in the development of autonomous agent capabilities for control of physical entities such as RPA's and their associated control surfaces. In FY13, the MIMFA program has two major Focus Areas of research, Collaborative Agency for Shared Awareness, and Classification and State Based Reasoning. Within each of these Focus Areas are several associated technical topic areas of interest. Our FY13 Focus Areas for MIMFA, with award details, are: Collaborative Agency for Shared Awareness: Decision makers can be expected to utilize numerous and diverse information gathering assets to gain a clear understanding of a situation prior to making decisions. These assets have the ability to sense and produce massive amounts of data, however, the volume of this data rapidly overwhelms the abilities of human operators to process it all, leading to potential delays in decision making or missed opportunities. This flood of data is further increased due to overlapping information collection duties of multiple assets that cause redundancies in the reported data. The vast quantity of available data can additionally have the negative effect of obscuring gaps in situational awareness (SA), which can lead to poorly informed decisions. Such delays and inaccuracies are unacceptable in dynamic and high-tempo situations. These challenges can be mitigated through self-optimizing information sharing capabilities both among the autonomous agents performing the tasks and between the agents and their operators. Autonomous information systems that are capable of self-optimizing, with regard to a set of objectives, and adapting to evolving conditions are key components to improving the efficiency, agility, and capabilities of our nation's military. Providing situational awareness (SA), a complete but concise description of all relevant information to a commander, is a function of Command and Control (C2) and an application area that will benefit greatly from autonomous systems. Future C2 systems will depend on multiple heterogeneous autonomous systems/agents working together to complete shared and individual tasks that achieve and maintain situational awareness (SA). How these agents will collaborate (agent roles, tasking and re-tasking policies, etc.) under dynamic and adverse conditions remains an open research question and drives this FY13 focus area. Topic areas of interest within the FY13 Collaborative Agency for Shared Awareness Focus Area include: Intelligent Information Sharing - Novel approaches are sought in the area of intelligent information sharing between autonomous agents to provide decision makers with the relevant knowledge needed to make a decision in time for it to have a positive impact on the mission. Several technical objectives of interest fall under this broad challenge and relate to the collection, processing, and dissemination of information amongst a heterogeneous system of collaborative agents in the face of changing objectives and environmental threats. For example, agents need the ability to mine knowledge from their own collected data and represent these findings in a way that can be comprehended by interested parties (as in knowledge transfer). Agents will also need to be able to identify what information should be propagated and to whom; both in pursuit of an agent's individual objectives and to further the objectives shared by all agents in the system. Finally, communication policies to enable information sharing must be developed to allow agents to complete mission objectives in dynamic and contested environments where accessible agents and lines of communication can change over time. The ability of a system of autonomous agents to rapidly identify, obtain, and relay necessary information to a decision maker will be key to revolutionize how military operations are conducted in the future. Autonomous Dynamic Coordination and Tasking - Novel approaches are sought for the coordination, organizational structures, and tasking of autonomous agents to provide superior SA in dynamic and adversarial environments. This topic area will address multiple challenges to achieve its objective: dynamic tasking and re-tasking protocols necessary for the optimal utilization of available resources. Scalable distributed learning algorithms are required to adapt the tasking behaviors of agents as situations evolve. Novel coordination protocols to enable timely and efficient dissemination of relevant information. Additionally, adaptive agent organizational structures that evolve and optimize with changing mission priorities and objectives are of interest. The overall objective or commander's intent shall be defined for the agents. Assumptions regarding knowledge sharing and processing protocols between agents must be clearly outlined. Individual agents may be stochastic and/or deterministic processes, but an aggregate system of agents must be bound by a clearly stated condition and be deterministic in terms of feedback, i.e. provide an answer when a certain condition is met. White papers, for these two topics only, will be due by: 12 Sep 2012. It is anticipated that awards for this particular topic will range from 9 to 24 months with dollar amounts nominally up to $500,000 per award. Please email your white papers to: Edward.Verenich@rl.af.mil Classification and State Based Reasoning: Our nation's information analysts need access to autonomous intelligent analytic tools that are capable of functioning on large and complex data sets in order to preserve the information superiority of our nation's military. Static data sets today are measuring in size from tera to petabyte ranges and can span several thousand features. Ultimately, autonomous agents will be expected to extract knowledge from several heterogeneous data sets of this magnitude in parallel. Additionally, these agents must function on data sets possessing dynamic qualities, both with respect to growing numbers of samples, varying features, and transient data collection and analysis systems. These autonomous agents will participate in the Planning, Collection, Processing, Analysis, and Dissemination (PCPAD) information flow framework to provide situational awareness (SA), a complete but concise description of all relevant information, to a commander. Of broad interest to this focus area are methods for robust and resilient machine learning (ML). ML is the capability for software to adapt through training or experience to better solve problems. These areas house a diverse set of techniques that can help address the challenges associated with extracting knowledge from complex, real-world data sources. Listed below are three topic areas of interest to the Classification and State Based Reasoning focus area. In each of these topic areas, data sources will be streaming data (either constantly or periodically) and proposed approaches are encouraged to address this. Topic Areas of interest within the FY13 Classification and State Based Reasoning Focus Area include: Problem Abstraction, Reduction, and Reformulation In many cases, available data is too complex for timely processing using data mining and machine learning. This is especially true in highly relational settings and setting with large feature spaces. As a result, conclusions must be drawn from manual inspection or using a significantly smaller subset of the data. This leads to decisions being made without full consideration of the available data. New methods are needed for state abstraction, dimensionality reduction, and problem reformulation, including automatic task decomposition, to compactly represent otherwise large problems and make them applicable to existing data mining and machine learning techniques. Several techniques of interest include, but are not limited to, the following: a) State abstraction; which promotes generalization in learning by sharing knowledge from one system state across all members that belong to the same abstract state. State abstraction approaches should be applicable in on-line learning scenarios and should allow for efficient re-computation as new information becomes available. b) Dimensionality reduction; a collection of techniques that seek to capture all relevant information about a data set and express it using a compact feature set. Proposed dimensionality reduction approaches may include both feature selection and extraction techniques and must address how to identify relevant features over both static and dynamic data sets. c) Problem reformulation; the process by which a complex situation is analyzed and re-expressed into more comprehensible parts to facilitate learning and data mining. Problem reformulation methods should be able to identify and exploit structure in the data or the problem itself either during a static preprocessing phase or in dynamic situations as new information becomes available. Machine Learning in Untrustworthy Environments Machine Learning techniques are emerging as a vital tool for discovering hidden information in data and adapting to complex environments. Increased reliance on these methods introduces challenges that may undermine the effectiveness of automated systems. These challenges include data sources which do not include necessary information (missing data), include inconsistent or noisy data (uncertain data), and data which was deliberately provided to sensors to mislead knowledge discovery agents (adversarial data). The presence of one or more of these data types may reduce the accuracy of outputs, confound the selection of an appropriate action, or evoke the undesirable behavior from an automated learning system. Novel new methods are sought for the hardening of Machine Learning algorithms against incomplete, inaccurate, and/or malicious training data. Desired capabilities include, but are not limited to: a) Support for training data with varied levels of priority, certainty and/or accuracy. These values may change dynamically and should allow for evaluation of different scenarios that cast suspicion on certain data types, data sources, or time intervals. b) Identifying the sufficient conditions and states under which algorithm output(s) will be significantly affected by missing, noisy, or malicious data. Additionally, leverage this information to maintain guaranteed levels of performance. c) Methods for identifying data that may represent a hidden but intentional attack against the algorithm through the injection of data with the purpose of maximizing classification error. Such attacks may originate from a single source or be carried out as a coordinated effort from multiple sources and may be found in static, dynamic, or streaming data. Automated Feature Pre-Processing Machine Learning approaches often require that data is annotated and relevant to the current objective. Annotation of new or unknown data sets can be difficult or infeasible in some situations, especially those where the data sets are dynamic and directed data collection is not possible. Ideally, new data would be pre-processed to identify the context and applicability to current learning objectives without the need for analysis and interpretation by humans. Novel approaches are sought to automate pre-processing of data sets for use in Machine Learning algorithms. Desired capabilities include, but are not limited to: a) Determining the feasibility of exploitation for the given data set. This may include recognizing content type(s) of the possible features, determining if the data set contains enough information to suitably train a given algorithm, or the relevance of the data to the training objectives. b) Identifying the algorithm or ensemble of algorithms that are most likely to be effective based on the features of the data set or environment. Given limited time and/or computational resources, it may be infeasible to process the data using many different approaches. c) Leveraging the outputs from the application of learning algorithms to aid in future selection of algorithms for training on unknown or uncertain features. Such meta-learning approaches should support local learning on a single platform and distributed sharing of learned patterns. White papers, for these three topics only, will be due by: 12 Sep 2012. It is anticipated that awards for this particular topic will range from 9 to 24 months with dollar amounts nominally up to $500,000 per award. Please email your white papers to: Michael.Corey@rl.af.mil No other changes have been made.
- Web Link
-
FBO.gov Permalink
(https://www.fbo.gov/spg/USAF/AFMC/AFRLRRS/BAA-10-01-RIKA/listing.html)
- Record
- SN02843906-W 20120818/120817001732-5af2dd35fe600a1c7820df28013bb0df (fbodaily.com)
- Source
-
FedBizOpps Link to This Notice
(may not be valid after Archive Date)
| FSG Index | This Issue's Index | Today's FBO Daily Index Page |