LGP

Linear genetic programming (LGP) is a form of genetic programming in which candidate programs are represented as linear sequences of instructions that are executed sequentially. Each program operates on a fixed set of registers and uses a defined repertoire of primitives (operators) and terminals (constants and input variables). The final output is read from a designated register after the instruction sequence completes. LGP is an alternative to tree-based GP, often used for numeric and procedural tasks.

Representation and execution: A genome is a sequence of instructions. An instruction has an opcode and operands that reference registers or constants. During execution, operands read values from registers and results are stored back into registers. Some variants support conditional instructions or loops, but many implementations run the sequence in a straightforward, single pass and produce an output from a specific register.

Genetic operators and evolution: Evolution proceeds with a population of such genomes evaluated on a domain task. Operators include mutation (altering an instruction, its operands, or constants) and crossover (exchanging instruction blocks between parents). Selection methods such as tournament selection guide the search, often with elitism. LGP emphasizes efficient evaluation and scalable representations, with attention to register count and instruction set design to avoid inefficiency or overfitting.

Applications and history: Since its emergence in the 1990s as an alternative to tree-based GP, LGP has been applied to symbolic regression, time-series prediction, control, and other numeric tasks. It can yield compact, learnable programs that map inputs to outputs and may be directly compiled to imperative code. The approach remains a useful option in evolutionary computation research.