# Ashley Reid-Montanaro benchmarks

A CGPS is up to 2000 times faster than a GPS using Lisp - Peter Nordin and Wolfgang Banzhaf (1995). Evolving Turing-Complete Programs for a Register Machine with Self-modifying Code. In Genetic Algorithms: Proceedings of the Sixth International Conference (ICGA95)

## [edit] Intel CGPS vs. Koza Lisp

From Ashley Reid-Montanaro: A Compiling Genetic Programming System for the Intel Architecture

- 5-parity problem - 24 times faster than Koza’s Lisp system

### [edit] Koza

- 32 different binary numbers of length 5 as fitness cases for his GP system
- the logical functions (AND, OR, NAND, NOR)
- With a population size of 96,000, a correct solution to the problem is evolved in 13 generations
- 6,481 and 36,950 seconds
- the function evolved is extremely complex, using 160 Lisp terminals
- This function also makes extensive use of automatically defined functions (ADFs)

### [edit] Intel CGPS

- the standard function set, consisting of all possible arithmetic and logical operations
- a population size of 100,000, and a maximum program size of 10 instructions
- a correct solution to the problem was found in 17 generations
- 262 seconds on a Pentium III machine
- The evolved solution was 9 instructions long and used two of the processor’s registers
- A solution which only used one register was also evolved during another run, but required more space (20 instructions)