Tests as Epistemic Updates: Pursuit of Knowledge

32 Acknowledgements Thanks are due to Frans Voorbraak for providing us with example 3.12. 31 (2) after the agent performs the action s/he should know that s/he had just performed the action 5. (3) the agent knows that the observable result holds after the test is performed just in case that the observable precondition was true before the test was performed. These demands diier substantially from the demands that we formulated in section 2. The demands for truthfulness and minimal change are not implied by Moore's properties. Informativeness holds to the extent that in some successor state knowledge is acquired concerning the state in which the test was performed. However due to the fact that neither truthfulness nor minimal change is demanded it needs not be the case that this knowledge is still of use in the successor state. In the concrete example that is given in Moo84], viz. that of a test for the acidity of a solution using blue litmus paper, a demand for truthfulness is made in order to ensure that the knowledge of the agent still applies. The test that is used in this example is a very speciic one (a litmus test) and not an instance of a general one like the tests in our formalization. Summarizing, due to the diierent emphasis that is placed in Moo84], the tests considered there diier substantially from tests as we see them. In particular Moore's notion of tests cannot be used to formalize the idea of tests as observations as we introduced it in this paper. 6. Discussion In this paper we looked into the nature of tests as a means to acquire knowledge. We gave three demands that actions should meet for them to be called tests: informativeness, truthfulness, and the demand for minimal change. Furthermore some additional postulates are deened that are characteristic for tests performed in typical AI environments. A possible formalization of a general test action test ' is given. This formalization incorporates some elements of update semantics into the framework for events and abilities that we deened in HLM93a, HLM93b]. Two implementations of tests are given, one in a KT and one in an S5 setting. For both implementations it holds that the test action satisses both the general demands for tests and the additional demands imposed on AI tests. Furthermore, both the implementation of tests as we give them are such that …