This paper describes a hybrid neural network and rule-based pattern recognition system architecture which is capable of self-modification or learning. The central research issue to be addressed for a multiclassifier hybrid system is whether such a system can perform better than the two classifiers taken by themselves. The hybrid system employs a hierarchical architecture, and it can be interfaced with one or more sensors. Feature extraction routines operating on raw sensor data produce feature vectors which serve as inputs to neural network classifiers at the next level in the hierarchy. These low-level neural networks are trained to provide further discrimination of the sensor data. A set of feature vectors is formed from a concatenation of information from the feature extraction routines and the low-level neural network results. A rule-based classifier system uses this feature set to determine if certain expected environmental states, conditions, or objects are present in the sensors’ current data stream. The rule-based system has been given an a priori set of models of the expected environmental states, conditions, or objects which it is expected to identify. These models are represented as a directed graph of features vectors. The rule-based system forms many candidate directed graphs of various combinations of incoming features vectors, and it uses a suitably chosen metric to measure the similarity between candidate and model directed graphs. The rule-based system must decide if there is a match between one of the candidate graphs and a model graph. If a match is found, then the rule-based system invokes a routine to create and train a new high-level neural network from the appropriate feature vector data to recognize when this model state is present in future sensor data streams. A different high-level neural network is created for each model state identified by the rule-based system. Both the high-level neural networks and the rule-based system receive all subsequent feature vector data streams, and each of these classifiers provides an estimate of whether a certain model states exist in the data. If the high-level neural network finds that a model state is present and the rule-based system does not, then the rule-based system will alter its model of the state - the rule-based system learns from the high-level neural net. If the high-level neural network does not find that a model state is present and the rule-based system does, then the rule-based system will invoke a neural network retraining routine - the high-level neural network learns from the rule-base. A second research issue is whether the self-modification of models can be controlled, such that models are only altered to reflect new information and not erroneous information from noisy sensors.