Fast Learning of Human Interaction Behavior

Learning Patterns of User Interaction on the LEA Robot

More Info
expand_more

Abstract

This thesis describes three different predictive frameworks which are applied to improve the interaction behavior between a robot and their users. Machine learning is used to train the frameworks. The training data contains information obtained from one user over a period of approximately 3 to 30 days. This work is applied to the LEA robot: an elderly assistant robot.

Each framework is implemented in an exciting real problem with the aim to improve life quality of elderly. The first framework predicts when and from where the user calls the robot to go towards the user on a daily basis. The repetitive events are predicted with an accuracy of 98%. The framework starts to give correct predictions from four examples only. This framework uses a density-based clustering algorithm in combination with a shifting window over the input data. This combination creates a predictive framework that adapts to the user behavior changes.

The second framework is used to filter undesired warning messages such as e.g. obstacle warnings while walking. The true positive rate is measured to be 100% for events that occur at least once in the three days. The algorithm creates a grid map which is used to memorize trajectories that lead to undesired warning messages.

The last framework predicts the user destination while walking. This is an important aid for patients suffering from dementia. The framework uses the trajectory and the starting time and location as input. A forward neural network is used to classify the room destination of the user. The network has two input layers. The first input layer uses a part of
the trajectory and the second input layer is an embedding layer which reduces the size of "metadata". The framework classifies 66.6 % of the time with a certainty above 85%. When only considering predictions with a classifications certainty above 85%, a predictive accuracy of 97.3% is measured. If all classifications are considered, the framework showed to be 86.4 % accurate, whereas an alternative approach like support vector machines showed to have a predictive accuracy of 60 % on this problem.