Self-Modification of System Control Schemata

From AGIRI.org

The most obvious route to enabling Strong Self-Modification in the Novamente Cognition Engine is to apply the system's capability for Procedure Learning to the problem of learning procedures for controlling the system's MindAgents, as well as for carrying out more basic control processes like controlling MindAgent scheduling.

[edit]

Temporary, Pragmatic Obstacles

The current Novamente codebase does not support this in an optimal way, because the code inside the various MindAgents is C++ code. However, it would be quite possible to replace this C++ code with code in the Combo language, which could then be more easily comprehended, modified and/or replaced by the Novamente system's learning algorithms. The reason this has not been done so far is that:

the current Novamente instances are nowhere near smart enough to mess with their own control code
Combo code currently executes much more slowly than analogous C++ code, and the code in many of the MindAgents is performance-critical

The second problem is solvable given a large amount of relatively well-understood software engineering effort: what is needed is simply to make a Combo compiler that produces fast code for the programs of interest (which are mostly fairly small programs combining mathematical manipulations with access to the AtomTable). Right now Combo is interpreted, not compiled.

[edit]