Argumentation-Based Online Incremental Learning

The environment around general-purpose service robots has a dynamic nature. Accordingly, even the robot’s programmer cannot predict all possible external failures which robot may confront. This research proposes an online incremental learning method that can be further used to autonomously handle originating from change in environment. Existing typically offers special-purpose solutions. Furthermore, current algorithms generalize well with just few observations. In contrast, our extracts set of hypotheses, then for finding best recovery behavior at each failure state. proposed argumentation-based approach uses abstract and bipolar argumentation framework extract most relevant hypotheses model defeasibility relation between them. leads novel overcomes addressed problems different domains including robotic applications. We have compared state-of-the-art approaches, approximation-based reinforcement method, several contextual bandit algorithms. experimental results show learns more quickly lower number observations also higher final precision than other methods. Note Practitioners—This work faster by using states approaches. resulting technique Argumentation-based generates explainable rules human-robot interaction. Moreover, testing publicly available dataset suggests wider applicability outside robotics field wherever learner is required. limitation it aims handling discrete feature values.