LDST
An automatic tuning method that uses large language models to extract diagnostic information from solver logs, propose sparse configuration modifications based on parameter semantics, and validate and execute them, serving as an alternative to conventional black-box optimization.
Timeline (last 90 days)
The researchers propose log-driven semantic tuning (LDST), which uses large language models (LLMs) to automate the configuration tuning process of MIP solvers.
LDST extracts compact diagnostic information from solver logs, and based on parameter meanings and observed solver behavior, the LLM proposes sparse modifications to the current best configuration, with each proposal validated before execution.
LDST introduces a mechanism-adaptive objective: for solved runs it uses solve time, while for unsolved runs it estimates completion time from the branch-and-bound tree growth records in the logs.
On 215 instances from five benchmark sets, compared with the best performance among SMAC3, RBFOpt and random search, LDST reduced the geometric mean loss by 43.9% on easy instances and 16.7% on medium instances.
In both difficulty ranges, LDST reached with only 10 proposals a lower loss than the baseline methods reached with 50 proposals.
LDST configurations tuned on a single instance outperformed baseline configurations when transferred to other unseen instances in the same family.
LDST still depends on the reasoning quality of the LLM, and the way log diagnostics are extracted may affect the final results.
All Facts (7)
The researchers propose log-driven semantic tuning (LDST), which uses large language models (LLMs) to automate the configuration tuning process of MIP solvers.
50%UnverifiedLDST extracts compact diagnostic information from solver logs, and based on parameter meanings and observed solver behavior, the LLM proposes sparse modifications to the current best configuration, with each proposal validated before execution.
50%UnverifiedLDST introduces a mechanism-adaptive objective: for solved runs it uses solve time, while for unsolved runs it estimates completion time from the branch-and-bound tree growth records in the logs.
50%UnverifiedOn 215 instances from five benchmark sets, compared with the best performance among SMAC3, RBFOpt and random search, LDST reduced the geometric mean loss by 43.9% on easy instances and 16.7% on medium instances.
50%UnverifiedIn both difficulty ranges, LDST reached with only 10 proposals a lower loss than the baseline methods reached with 50 proposals.
50%UnverifiedLDST configurations tuned on a single instance outperformed baseline configurations when transferred to other unseen instances in the same family.
50%UnverifiedLDST still depends on the reasoning quality of the LLM, and the way log diagnostics are extracted may affect the final results.
50%