A Backtracking Strategy for Order-Independent Incremental Learning

Agents that exist in an environment that changes over time, and are able to take into account the temporal nature of experience, are commonly called incremental learners. It is widely known that incremental learning systems suffer from order effects, a phenomenon observed when differently ordered sequences of examples lead to different results. The goal of this paper is presenting INTHELEXback, an order-independent evolution of the incremental learning system INTHELEX. A backtracking strategy is incorporated in its refinement operators, which causes a change in its refinement strategy and reflects the human behavior during the learning process. It consists in remembering the different versions of the learned theory across modifications due to new evidence. In this way the system can backtrack on a previous knowledge level when it discovers to have made a wrong choice. Experiments on an artificial dataset validate the approach in terms of computational cost and predictive accuracy.