Learned Anticipation Strategy on Complex Behaviors and as an Approach to Generalization Behavior for the Coordination of an AUV Fleet

Abstract

Anticipation is natural behavior seen in certain organisms when they try to adapt their behavior to create an accurate reaction to an event that will happen in the future based on sequential information that extracted from the environment and other organisms. Anticipation has been studied from both the human and the animal perspectives and science has been taking some of these ideas to implement this behavior in machines. Research and studies of anticipation using Artificial Intelligence (AI) methods has been growing during the past years within different fields, including Robotics, Stock Market, Weather Forecast, Social Media, among others. At the University of Idaho, we have been focusing on studying Anticipation using Autonomous Underwater Vehicles (AUV). This paper analyzes the impact on coordination of including an anticipation module to adapt to complex behaviors and to make decisions based on generalization while running as part of a simulation of a fleet of AUVs following a magnetic signature assessment (MSA) task. During this task, a moving Target Ship (TS) shares information about its behavior with the fleet of AUVs while the fleet tries to maintain formation and reach a designated meeting point at the same time as the TS. The TS behaviors include sudden and/or complex changes in movement during the experiment. These changes in the TS’s movement make it difficult for the AUVs to reach the meeting point at the same time as the TS without additional information. Progress messages from the TS are used by the AUVs to adjust their behavior, but the messages may be sparse. A portion of the data that represents the TS’s behavior was used as training data for the anticipation module in the AUVs. The remaining data was used as brand new cases that the anticipation module would try to solve to encourage generalization. Generalization occurs when the robot can solve a task that it was not trained for, but there was sufficient training data to result in reasonable solutions. For these experiments, the messages that are sent by the TS to the AUVs report the TS’s progress to the goal. The TS reports its progress using Fuzzy set membership values. This strategy is used to model real-world situations where progress cannot be defined with exact values. The membership values are represented by fuzzy sets including On Schedule, Behind Schedule, Ahead of Schedule, etc. This strategy gives a more human-like approach to the AUVs for decision making because humans often say things like “I’m a little behind schedule”. The anticipation module works as an aid for the communication between the TS and the fleet of AUVs. It tries to keep the fleet running on track and synchronized with the TS behavior by filling in missing messages with an anticipated message. In this research it is assumed that missing and corrupt messages occur because of a noisy or low-bandwidth communications. The experiments simulate this condition by sending messages more infrequently. The TS stops broadcasting messages for a specific time period that has been selected before starting the test. Versions of the anticipation module, using a Neural Network model and a Fuzzy Logic model were tested. To have comparable results, the position of the meeting point is constant during each experiment and is used as a point of reference. The effectiveness of the anticipation module was evaluated by measuring the distance between the actual meeting point of the fleet of AUVs and the TS and the desired meeting point. In general the anticipation module reduced the error significantly. This included the tests using the novel behaviors that were not part of the training set. These last results lead to the idea that, with the right training data, the AUVs are displaying generalization behavior. In general, the anticipation module helps the fleet of AUVs to infer the behavior of the TS and they synchronize to the TS even when there was a lack of information, caused by missing messages.