NASA NTRS1989
PRIME is a system to be used by an intelligent machine to allow it to operate in an abstract but uncertain (or stochastic) environment. It maintains a model of the effects of the machine's actions in the form of a rule base, which is induced from experience. This bottom-up approach to rule development allows the model to adapt to changes in the environment. Each rule consists of a condition under which the rule is active, an action, the effect of the action on the environment, and an estimate of the probability of this effect occurring. The effect probabilities are used to model the uncertainty in the environment, permitting multiple possible effects for a single action under a particular set of conditions. The objective of the intelligent machine is to satisfy user-specified goals with maximum probability of success. PRIME fulfills this requirement in two ways: it continuously updates the rule base with the most recent information, to ensure the validity of the model; and it generates plans which have the maximum probability of achieving the goals, based on the probability estimates in the rule base. PRIME is composed of three main processes: exploration, generalization, and planning. In exploration, the machine executes various randomly chosen actions, observes the effects on the environment, and updates the rule base accordingly. This process is used to develop the rule base in simulation, as well as to supplement the current knowledge during normal operation. Generalization is the procedure used to induce general rules from experience, which is encoded in the form of specific rules. These general rules extend the machine's knowledge to situations which have not been encountered yet, thereby increasing the capability of the machine to plan effectively. Planning is the process of constructing an optimal sequence of actions to satisfy a goal, using the rule base to predict the effects of these actions and to determine the probability of success of the plan. The rule representation and many other data structures were specifically chosen to maximize the efficiency of these processes. A simulated environment was designed to test the performance of PRIME. The results of experimentation were largely negative. The main problem was that the domain coverage of the rules was inadequate for the number of rules stored in the rule base, due to redundancies in general rules and numerous rules covering ineffective actions. It was determined that a more efficient generalization, and some form of goal-directed exploration, are necessary in order to solve most of the current deficiencies in PRIME.