Despite extensive research on innovation subsidies, the critical question of how to optimally assign these public subsidies to firms remains largely unexplored. This paper introduces an optimal policy learning (OPL) approach to map firm characteristics to subsidy assignment under welfare maximization of firm-innovation output, using (i) threshold-based, (ii) linear-combination, and (iii) fixed-depth decision tree policies over a large sample of Spanish firms. Exploiting comprehensive data from the Spanish Technological Innovation Survey, I construct an aggregate measure of innovation output using a generalized least squares weighting procedure, relying on a standardized inverse-covariance weighted average of innovation outcomes. Harnessing a mix of subsidies managed at national, regional, and European levels, critical firm characteristics are selected from a large pool of variables on the basis of a post-double-selection Lasso method. Leveraging regression adjustment and causal forests to estimate the heterogeneous treatment effects of subsidies on innovation, OPL findings indicate that innovation expenditures and firm size consistently emerge as key determinants for subsidy allocation, with decision trees providing maximal expected constrained welfare. The paper offers actionable insights for policymakers, emphasizing tailored subsidy frameworks over generic, one-size-fits-all approaches. The proposed approach not only optimizes innovation output but also provides a scalable and replicable framework for subsidy allocation, within institutional contexts similar to Spain’s multi-level innovation subsidy system.
This study addresses the challenges of wind speed intermittency in wind power generation, aiming to enhance wind speed modeling for improved energy efficiency and reliability. We systematically refined each step of the modeling process, beginning with an exploration of 14 PDFs, including Weibull, Gamma, their length-biased variants, and bimodal mixtures. A grid search algorithm was integrated with five parameter estimation methods to optimize bimodal mixtures given their complexity. A total of 70 models were evaluated across twelve sites with diverse climatic and topographical characteristics. To assess the trade-off between wind speed modeling accuracy and wind power density estimation, various goodness-of-fit metrics were employed. Additionally, the k-means clustering, and Principal Component Analysis were used to identify the most suitable PDFs for specific site characteristics, addressing a key gap in wind power assessment literature. Key findings highlight the superior performance of length-biased models. Class A, which includes three bimodal length-biased mixtures achieved the best balance between wind speed modeling and wind power density estimation. Notably, bimodal length-biased Weibull exhibited consistent performance across all estimation methods. The results also suggest potential biases in data collection process. To our knowledge, length-biased mixture distributions have not been previously explored in wind power modeling.
Do International Monetary Fund (IMF) programs foster or hinder economic development in Africa? This paper addresses this long-standing and contested question by estimating the causal impact of IMF program participation on real GDP per capita across African countries from 1990 to 2021. Departing from previous studies that often rely on strong parametric assumptions or limited controls, I leverage a double/debiased machine learning framework that accommodates high-dimensional confounding and flexible functional forms. By combining rigorous causal inference with rich historical and structural data, this study provides new evidence on the developmental consequences of international financial intervention, while accounting for impact heterogeneity. Using data on the Fund’s three major instruments–the Extended Fund Facility, the Stand-By Agreement, and the Extended Credit Facility (ECF)–I find that program participation is associated with large and persistent declines in economic development. On average, IMF-supported programs reduce real GDP per capita by over 50%, with the ECF exhibiting the most severe impact. These findings are robust to alternative estimators, and hold across a range of sensitivity and falsification tests. Causal forest estimates further reveal significant heterogeneity in treatment effects, with severe impacts concentrated in fragile states with low human development. Importantly, the near-constant detrimental effects across decades highlight the persistent structural rigidities in the IMF program design. These results suggest that adverse development effects are likely to materialize through several mechanisms and call for a critical reassessment of the IMF program alignment with long-term development priorities in Africa.
This study evaluated the growth kinetics and hyoscyamine biosynthesis of transgenic hairy roots from Datura innoxia (DI), Datura stramonium var. stramonium (DS) and Datura stramonium var. tatula (DT) induced by Agrobacterium rhizogenes strain A4 and cultured in vitro for five years after their induction using liquid chromatography–tandem mass spectrometry for alkaloid analysis. The lines were elicited with AgNO3 for 24 and 48 h to determine its effects on the biomass, hyoscyamine and scopolamine yield of hairy roots. DT grew better, resulting in 0.295 g of dry mass after 21 d of culture on B5 medium in the first year, compared to 0.229 g for DI and 0.205 g for DS. Over five years, DT grew better than DS and DI. Furthermore, the hyoscyamine yield was higher in DT than in DI and DS with the best yield occurring in the first year of culture (8.575 mg g−1 DW). After chemical elicitation, the biomass, expressed as the dry weight of DS and DT, decreased under all AgNO3 treatments. However, the DI biomass increased after AgNO3 treatment for 24 and 48 h (74%) of the control. Only two treatments promoted hyoscyamine biosynthesis, and the best yield was obtained with 7.89 mg g−1 DW of hyoscyamine after 48 h of elicitation. However, the scopolamine yield was improved by AgNO3 elicitation for 24 h with the best result obtained for DT hairy roots (0.539 mg g−1 DW).
Vision Transformers allocate most parameters to multi-layer perceptrons (MLPs) for channel mixing, while token interactions usually rely on quadratic multi-head self-attention (MHSA). Linear attention reduces sequence complexity to O(N), but remains coupled with the same fixed-activation MLP as softmax Transformers. Kolmogorov-Arnold Networks (KANs) instead place learnable univariate maps on edges, yet prior vision KANs keep MHSA or omit attention entirely. We introduce LKAT (Linear Kolmogorov-Arnold Transformer), an isotropic ViT encoder that couples chunkwise Gated Linear Attention (GLA) with a two-layer KAN feed-forward, and we provide an I/O-aware fused RBF-KAN kernel for the radial-basis grid maps. Under a shared DeiT-style recipe we compare LKAT with ViT, ViT-5, and MLP-Mixer. LKAT-B exceeds ViT-B/16, ViT-5-B, and Mixer-B/16 on ImageNet-100. Tiny/Small/Base LKAT variants scale consistently on CIFAR-10/100, and ImageNet-100 pretraining transfers to CIFAR fine-tuning. The results support gated linear attention and KAN-based radial-basis functions as complementary inductive biases for mid-scale visual representation learning.
