On the Impossibility of Global Convergence in Multi-Loss Optimization
Authors: Alistair Letcher
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | We negatively resolve this open problem by proving that desirable convergence properties cannot simultaneously hold for any algorithm. Our result has more to do with the existence of games with no satisfactory outcomes, than with algorithms per se. More explicitly we construct a two-player game with zero-sum interactions whose losses are both coercive and analytic, but whose only simultaneous critical point is a strict maximum. |
| Researcher Affiliation | Academia | The paper lists the author as 'Alistair Letcher aletcher.github.io'. No specific institutional affiliation (university or company name) is explicitly provided in the paper's header or body, preventing a definitive classification. |
| Pseudocode | No | The paper describes algorithms mathematically in Appendix A but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Accompanying code for all experiments can be found at https://github.com/aletcher/ impossibility-global-convergence. |
| Open Datasets | No | The paper uses mathematically constructed 'games' (M and N) for its theoretical proofs and simulations, rather than publicly available datasets. |
| Dataset Splits | No | The paper does not specify training/validation/test dataset splits, as its focus is on theoretical proofs and simulations of constructed games. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types) used for running its simulations. |
| Software Dependencies | No | The paper mentions 'Singular' for a mathematical proof but does not provide specific version numbers for any key software components or libraries used for the general experimental setup. |
| Experiment Setup | Yes | In all experiments we initialise θ0 following a standard normal distribution and use a learning rate α = 0.01, with γ = 0.01 for CO. |