Conflict Resolution in Advice Taking and Instruction for Learning Agents

We raise and discuss several issues of advice taking and instruction, including: the challenges of very small samples of opinion, and of assimilation with a large, complex prior knowledge base. We focus expecially on the problem of conflicting yes/no rules. We observe the availability of natural kinds of information, e.g., authority, reliability, freshness, and specificity, as teh basis for reasoning about precedence. by precedence we mean in the sense of resolving conflicts between rules on the basis of qualitative ordinal information. We propose an approach to this problem of conflict: via defaults and reasoning about such precendence. drewn from knowledge representation and commonsense reasoning. We initially developed this approach in previous work. Here, we elaborate it, abstract it to a more conceptual level, and present it for a machine learning audience. Also, we abstract it away from the details of, and dependency on, one aprticular non-monotonic logical formalism. We see the value of this approach not as being all-embracing, but as providing a step towards elements of future approaches to advice-taking and instruction.