A new run-time dynamic optimization to reduce the upper bound of real-time rule-based expert systems is presented. It constructs a predicate dependency list, which consists of a predicate, active rule set, and inactive rule set, for each predicate in a real-time rule-based program. It then generates an inactive rule set by combining all false rules in a predicate dependency list together, and dynamically selects rules to be evaluated at run-time based on the inactive rule set. For the timing analysis of the proposed algorithm, we introduce a predicate-based rule dependency graph and its construction algorithm. We also discuss the bounded time for the EQL rule-based program using the predicate-based rule dependency graph. The performance evaluation shows that the dynamic optimizer reduces the number of rule evaluations and the number of predicate evaluations as well as the response time upper bound significantly, and the new algorithm has better upper bound comparing to the other optimization methods.