Action Prediction Using a Mental-Level Model

Weproposea formalapproachto the problem of prediction basedon the followingsteps: First, a mental-level modelis constructedbasedon the agent's previousactions; Consequently, the modelis updatedto accountfor anynewobservationsby the agent, andfinally, wepredict the optimalaction w.r.t, the agent's mentalstate as its next action. Thispaperformalizesthis prediction process. In orderto carry out this process weneedto understandhowa mentalstate call be ascribed to an agent, andhowthis mental state should be updated. In [Brafmanand Tennenholtz, 1994b] we examined the first stage. Here we investigate a particular update operator, whoseuse requires makingonly weakmodeling assumptions, andshowthat this operatorhas a number of desirable properties. Finally, we provide analgorithm for ascribingthe agent's mentalstate underthis operator.