Home     Technology   Products  & Services   About
 
 
   
       

Collaborative Learning Agent (CLA)  

 

 Summary

A single Collaborative Learning Agent learns and discovers knowledge and behavior patterns from historical data and then applies the patterns for identification of patterns in the new data. The knowledge patterns, discovered automatically using machine learning and pattern recognition methods, include the following patterns 1) Similarity patterns, i.e. group similar data; 2) Correlation patterns, i.e. find hidden relationships among data; 3) Predictive patterns, i.e. make predictions based on historical data; and  4) Recommendation patterns, i.e. make predictions when little or no historical data is available. 

A set of networked Collaborative Learning Agents (CLAs) forms an agent network include the following capabilities 1) Text mining: extract concepts and meaning clusters based on contexts; 2) Machine learning: extract knowledge patterns that link meaning to raw text or data observation; and 3) Collaborative meaning search: incorporate human and machine a single loop to form a collaborative network to search and enhance the meaning iteratively. 

A text mining technique, Context-Concept-Cluster (CCC) model (US Patent Pending), is implemented in the CLAs. The advantage of such a text mining technique over the traditional information retrieval is to capture the cognitive level of understanding of text observations.  

Machine learning starts with supervised learning. A train data set with both observations and their labeled meaning are presented to the learning system. The system then generates a correlation model between the categories of observations and meaning labeled by human analysts.  In real-life, however, human labeled meaning is expensive to obtain, therefore, it is more important to develop unsupervised learning to achieve the same goal. Here we want to show that CLAs perform an unsupervised learning and categorize a new information into four categories. 1) Anomaly category showing a search input  has low correlation with previously discovered context patterns; 2) Relevant category showing an input is highly correlated to the previously discovered knowledge patterns; 3) Low: between relevant and anomaly; 4) Irrelevant category showing an input is not related to any of the agents. 

The advantage of using machine learning is for automation and reducing manpower. For example, to prevent potential threats and events of terrorism related to the global maritime domain, DOD, federal, state and local agencies have to work together to make situation awareness effectively based on a Common Operational Picture (COP). In this environment there are numerous trouble reports.  These trouble reports are a usually assessed, analyzed and resolved by human analysts. Agent machine learning provides a mechanism to transfer knowledge patterns of human analysts to computer agents. Therefore manpower required for handling the trouble reports can be greatly reduced. 

In Trident Warrior 08, which is an annual FORCEnet SEA Trial experiment, we used three agents as shown in Figure 1 in this exercise to apply anomaly detection in this area from unstructured data sources. For example, three agents can be MDA partners representing inter-agency collaborations among navy, police, coast guard.  Then we applied a single CLA to each data source to discover context patterns representing normal behavior patterns.  The discovered knowledge patterns serve as the baselines or normal profiles for new data to compare with. The knowledge patterns are stored as a search index. When a piece of real-time information is newly observed, it goes through the search network of CLAs and is classified into categories of anomaly, relevant, median correlation or irrelevant. 

Figure 1: Trident Warrior Architecture

 
         
         
Site Map    |    Privacy    |    Terms  
Copyright 2001-2008 Quantum Intelligence Inc. All rights reserved.
 

setstats