Rule selection in forward chaining is a critical factor in the performance of expert systems. Uninformed selection causes many rules to be fired, that are not useful in the attainment of the reasoning goal. As a result, users have to answer more questions than needed and the system's performance is degraded. Domain-specific meta-rules have been used to improve rule selection. However, in most instances, such meta-rules are elicited from experts and thus costly to obtain and difficult to validate. This paper describes a method for improving rule selection automatically and adaptively. The idea rests on the use of a neural network to (meta-)learn dynamically (i.e., whilst the expert system is run) in which situations which rules are worth applying. Prior meta-rules, if they exist, may also be improved by the neural network. Empirical results demonstrate that the (meta-)knowledge encoded in the neural network produces a reduction of the number of rules selected during forward chaining.