| Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. |
Variante : Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. |
| Nom : | Esteban | |||
| Marque : | Fabricant : | |||
| Publicité : | ||||
| Pays : | Esteban | |||
| Date de début : | 0 | Date de fin : | 0 | |
| Type : | () | |||
| Dimensions, fermé : | Ouvert : | |||
| Poids : | ||||
| Matière : | Couleur : | |||
| Surf. sensible : | Format : | |||
Objectif : | ||||
| - Nom : | Esteban | Type : | ||
| - Marque : | ||||
| - Ouverture : | Focale : | |||
| - Fixation : | XXX | Monture : | XXX | |
Diaphragme : | ||||
| - Type : | XXX | |||
| - Ouverture : | x - y - z ou x à z | |||
Obturateur : | ||||
| - Nom : | Esteban | Type : | ||
| - Marque : | ||||
| - Vitesses : | x s - y s - z s ou x s à y s | Pose : | XXX | |
| - Retardateur : | XXX | |||
| - Armement : | XXX | |||
Film, plaque : | ||||
| - Compte-vues 1 : | XXX | Compte-vues 2 : | XXX | |
| - Protection 2 expos : | XXX | |||
| - Avancement : | XXX | |||
Flash : | ||||
| - Synchro : | XXX | |||
| - Flash interne : | XXX | |||
Posemètre : | ||||
| - Type : | autre (12) | |||
| - Plage : | 894723083 | |||
| - Automatisme : | manuel | |||
Mise au point : | ||||
| - Type : | XXX | Minimum : | ||
| - Télémètre : | réflex TTL (SLR) | |||
Viseurs : | ||||
| - Viseur 1 : | XXX | |||
| - Viseur 2 : | XXX | |||
| - Viseur 3 : | XXX | |||
Autres : | ||||
| Niveau à bulle : | XXX | |||
| Porte-acc. 1 : | sabot synchronisé | Porte-acc. 2 : | autre (20) | |
| Écrou de pied : | XXX | |||
| Déc. vertical : | Horizontal : | |||
| Bascule avant : | Arrière : |
| Commentaires : | Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. | |||
| Liens web : | ||||
| https://98.staikudrik.com/index/d1?diff=0&utm_clickid=uskkokskw44sooos&aurl=http%3A%2F%2Fspybot.io | ||||
Appareil présenté : | ||||
| N° de série : | ||||
| N° objectif : | N° obturateur : | |||
| Accessoires : | Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. | |||
| Notice : | ||||
| Commentaires : | Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. | |||
| N° inventaire : | ||||
| descr_nok, , , sh_XXX | ||||
| État : | Wanneroo | |||
| Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. | ||||
| Achat : | ||||
| Vente : | ||||
| Estimation : | ||||
| Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. | ||||
| Pub, livres : | Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back eva | |||
| Commentaires : | Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. | |||
| Ancien texte : | Abstract:Expensive optimization issues (EOPs) are black-field tasks with costly objective evaluations and no gradient access, making the evaluation finances the important thing bottleneck. Surrogate-assisted evolutionary algorithms (SAEAs) cut back evaluations via surrogate predictions, but standard surrogates typically require frequent retraining as populations evolve, incurring overhead. This paper proposes R2SAEA, a reinforcement-educated relation-based large language model (LLM) surrogate assisted evolutionary algorithm. We forged relation-primarily based surrogate modeling as an in-context pairwise reasoning process. To allow environment friendly inference in evolutionary loops, we develop an anchor-primarily based iterative context development technique that reduces immediate complexity from quadratic to linear in inhabitants size, and a voting-primarily based aggregation scheme that converts predicted relations into scores for offspring choice. We additional build an RL pipeline from evolutionary trajectories and nice-tune Qwen2.5 with GRPO. Experiments on single- and multi-goal benchmarks show improved relation prediction and state-of-the-artwork optimization performance over strong SAEA baselines and general LLMs. Quantization additionally allows environment friendly edge deployment, supporting a zero-shot surrogate paradigm with out per-generation retraining. Code and models are available at this https URL. | |||