Learning through Overcoming Inconsistencies

Of all the perspectives about what inconsistencies entail and how we can handle them, one that escapes our attention is that inconsistencies can serve as effective stimuli to learning because they often help reveal the inadequacies, gaps, deficiencies, or boundary conditions in an agent's problem-solving knowledge. In this paper, we describe a new machine learning approach: inconsistency-induced learning where a learning episode is triggered by some inconsistent phenomenon and learning is essentially embodied in the process of finding ways to overcome such inconsistent phenomenon. Each learning episode causes an agent's task-performing knowledge to be revised, refined, or augmented, which in turn incrementally improves its problem-solving performance. If such a learning agent can be engaged in an open-ended and alternating sequence of task-performing episodes and learning episodes, we can refer to it as a perpetual learning agent. We also provide an overview of related work in the paper.