Grandmaster level in StarCraft II using multi-agent reinforcement learning.

Many real-world applications require artificial agents to compete and coordinate with other agents in complex environments. As a stepping stone to this goal, the domain of StarCraft has emerged as an important challenge for artificial intelligence research, owing to its iconic and enduring status among the most difficult professional esports and its relevance to the real world in terms of its raw complexity and multi-agent challenges. Over the course of a decade and numerous competitions1-3, the strongest agents have simplified important aspects of the game, utilized superhuman capabilities, or employed hand-crafted sub-systems4. Despite these advantages, no previous agent has come close to matching the overall skill of top StarCraft players. We chose to address the challenge of StarCraft using general-purpose learning methods that are in principle applicable to other complex domains: a multi-agent reinforcement learning algorithm that uses data from both human and agent games within a diverse league of continually adapting strategies and counter-strategies, each represented by deep neural networks5,6. We evaluated our agent, AlphaStar, in the full game of StarCraft II, through a series of online games against human players. AlphaStar was rated at Grandmaster level for all three StarCraft races and above 99.8% of officially ranked human players.

Simulation Study of Radio Frequency Safety and the Optimal Size of a Single-Channel Surface Radio Frequency Coil for Mice at 9.4 T Magnetic Resonance Imaging.

The optimized size of a single-channel surface radio frequency (RF) coil for mouse body images in a 9.4 T magnetic resonance imaging (MRI) system was determined via electromagnetic-field analysis of the signal depth according to the size of a single-channel coil. The single-channel surface RF coils used in electromagnetic field simulations were configured to operate in transmission/reception mode at a frequency of 9.4 T-400 MHz. Computational analysis using the finite-difference time-domain method was used to assess the single-channel surface RF coil by comparing single-channel surface RF coils of varying sizes in terms of |B1|-, |B1+|-, |B1-|- and |E|-field distribution. RF safety for the prevention of burn injuries to small animals was assessed using an analysis of the specific absorption rate. A single-channel surface RF coil with a 20 mm diameter provided optimal B1-field distribution and RF safety, thus confirming that single-channel surface RF coils with ≥25 mm diameter could not provide typical B1-field distribution. A single-channel surface RF coil with a 20 mm diameter for mouse body imaging at 9.4 T MRI was recommended to preserve the characteristics of single-channel surface RF coils, and ensured that RF signals were applied correctly to the target point within RF safety guidelines.

A Comparative Study of Birdcage RF Coil Configurations for Ultra-High Field Magnetic Resonance Imaging.

Improvements in transmission and reception sensitivities of radiofrequency (RF) coils used in ultra-high field (UHF) magnetic resonance imaging (MRI) are needed to reduce specific absorption rates (SAR) and RF power deposition, albeit without applying high-power RF. Here, we propose a method to simultaneously improve transmission efficiency and reception sensitivity of a band-pass birdcage RF coil (BP-BC RF coil) by combining a multi-channel wireless RF element (MCWE) with a high permittivity material (HPM) in a 7.0 T MRI. Electromagnetic field (EM-field) simulations, performed using two types of phantoms, viz., a cylindrical phantom filled with oil and a human head model, were used to compare the effects of MCWE and HPM on BP-BC RF coils. EM-fields were calculated using the finite difference time-domain (FDTD) method and analyzed using Matlab software. Next, to improve RF transmission efficiency, we compared two HPM structures, namely, a hollow cylinder shape HPM (hcHPM) and segmented cylinder shape HPM (scHPM). The scHPM and MCWE model comprised 16 elements (16-rad BP-BC RF coil) and this coil configuration demonstrated superior RF transmission efficiency and reception sensitivity along with an acceptable SAR. We expect wider clinical application of this combination in 7.0 T MRIs, which were recently approved by the United States Food and Drug Administration.

A Preliminary Study for Reference RF Coil at 11.7 T MRI: Based on Electromagnetic Field Simulation of Hybrid-BC RF Coil According to Diameter and Length at 3.0, 7.0 and 11.7 T.

Magnetic resonance imaging (MRI) systems must undergo quantitative evaluation through daily and periodic performance assessments. In general, the reference or standard radiofrequency (RF) coils for these performance assessments of 1.5 to 7.0 T MRI systems have been low-pass-type birdcage (LP-BC) RF coils. However, LP-BC RF coils are inappropriate for use as reference RF coils because of their relatively lower magnetic field (B1-field) sensitivity than other types of BC RF coils, especially in ultrahigh-field (UHF) MRI systems above 3.0 T. Herein, we propose a hybrid-type BC (Hybrid-BC) RF coil as a reference RF coil with improved B1-field sensitivity in UHF MRI system and applied it to an 11.7 T MRI system. An electromagnetic field (EM-field) analysis on the Hybrid-BC RF coil was performed to provide the proper dimensions for its use as a reference RF coil. Commercial finite difference time-domain program was used in EM-field simulation, and home-made analysis programs were used in analysis. The optimal specifications of the proposed Hybrid-BC RF coils for them to qualify as reference RF coils are proposed based on their B1+-field sensitivity under unnormalized conditions, as well as by considering their B1+-field uniformity and RF safety under normalized conditions.

An End-to-End Recurrent Neural Network for Radial MR Image Reconstruction.

Recent advances in deep learning have contributed greatly to the field of parallel MR imaging, where a reduced amount of k-space data are acquired to accelerate imaging time. In our previous work, we have proposed a deep learning method to reconstruct MR images directly from k-space data acquired with Cartesian trajectories. However, MRI utilizes various non-Cartesian trajectories, such as radial trajectories, with various numbers of multi-channel RF coils according to the purpose of an MRI scan. Thus, it is important for a reconstruction network to efficiently unfold aliasing artifacts due to undersampling and to combine multi-channel k-space data into single-channel data. In this work, a neural network named 'ETER-net' is utilized to reconstruct an MR image directly from k-space data acquired with Cartesian and non-Cartesian trajectories and multi-channel RF coils. In the proposed image reconstruction network, the domain transform network converts k-space data into a rough image, which is then refined in the following network to reconstruct a final image. We also analyze loss functions including adversarial and perceptual losses to improve the network performance. For experiments, we acquired k-space data at a 3T MRI scanner with Cartesian and radial trajectories to show the learning mechanism of the direct mapping relationship between the k-space and the corresponding image by the proposed network and to demonstrate the practical applications. According to our experiments, the proposed method showed satisfactory performance in reconstructing images from undersampled single- or multi-channel k-space data with reduced image artifacts. In conclusion, the proposed method is a deep-learning-based MR reconstruction network, which can be used as a unified solution for parallel MRI, where k-space data are acquired with various scanning trajectories.

A 2D-GRAPPA Algorithm with a Boomerang Kernel for 3D MRI Data Accelerated along Two Phase-Encoding Directions.

For the reconstruction of 3D MRI data that are accelerated along the two phase-encoding directions, the 2D-generalized autocalibrating partially parallel acquisitions (GRAPPA) algorithm can be used to estimate the missing data in the k-space. We propose a new boomerang-shaped kernel based on theoretic and systemic analyses of the shape and dimensions of the kernel. The reconstruction efficiency of the 2D-GRAPPA algorithm with the proposed boomerang-shaped kernel (i.e., boomerang kernel (BK)-2D-GRAPPA) was compared with other 2D-GRAPPA algorithms that utilize different types of kernels (i.e., EX-2D-GRAPPA and SK-2D-GRAPPA) based on computer simulation, phantom and in vivo experiments. The proposed method was validated for different sets of ACS lines with acceleration factors from four to eight and various sizes of the kernels. A quantitative analysis was also performed by comparing the normalized root mean squared error (nRMSE) in the images and the undersampled edges. Computer simulation, in vivo and phantom experiments, and the quantitative analysis, showed that the proposed method could reduce aliasing artifacts without reducing the SNRs of the reconstructed images.

A New Combination of Radio-Frequency Coil Configurations Using High-Permittivity Materials and Inductively Coupled Structures for Ultrahigh-Field Magnetic Resonance Imaging.

In ultrahigh-field (UHF) magnetic resonance imaging (MRI) system, the RF power required to excite the nuclei of the target object increases. As the strength of the main magnetic field (B0 field) increases, the improvement of the RF transmit field (B1+ field) efficiency and receive field (B1- field) sensitivity of radio-frequency (RF) coils is essential to reduce their specific absorption rate and power deposition in UHF MRI. To address these problems, we previously proposed a method to simultaneously improve the B1+ field efficiency and B1- field sensitivity of 16-leg bandpass birdcage RF coils (BP-BC RF coils) by combining a multichannel wireless RF element (MCWE) and segmented cylindrical high-permittivity material (scHPM) comprising 16 elements in 7.0 T MRI. In this work, we further improved the performance of transmit/receive RF coils. A new combination of RF coil with wireless element and HPM was proposed by comparing the BP-BC RF coil with the MCWE and the scHPM proposed in the previous study and the multichannel RF coils with a birdcage RF coil-type wireless element (BCWE) and the scHPM proposed in this study. The proposed 16-ch RF coils with the BCWE and scHPM provided excellent B1+ field efficiency and B1- field sensitivity improvement.

Voxel-wise partial volume correction method for accurate estimation of tissue sodium concentration in 23 Na-MRI at 7 T.

Sodium is crucial for the maintenance of cell physiology, and its regulation of the sodium-potassium pump has implications for various neurological conditions. The distribution of sodium concentrations in tissue can be quantitatively evaluated by means of sodium MRI (23 Na-MRI). Despite its usefulness in diagnosing particular disease conditions, tissue sodium concentration (TSC) estimated from 23 Na-MRI can be strongly biased by partial volume effects (PVEs) that are induced by broad point spread functions (PSFs) as well as tissue fraction effects. In this work, we aimed to propose a robust voxel-wise partial volume correction (PVC) method for 23 Na-MRI. The method is based on a linear regression (LR) approach to correct for tissue fraction effects, but it utilizes a 3D kernel combined with a modified least trimmed square (3D-mLTS) method in order to minimize regression-induced inherent smoothing effects. We acquired 23 Na-MRI data with conventional Cartesian sampling at 7 T, and spill-over effects due to the PSF were considered prior to correcting for tissue fraction effects using 3D-mLTS. In the simulation, we found that the TSCs of gray matter (GM) and white matter (WM) were underestimated by 20% and 11% respectively without correcting tissue fraction effects, but the differences between ground truth and PVE-corrected data after the PVC using the 3D-mLTS method were only approximately 0.6% and 0.4% for GM and WM, respectively. The capability of the 3D-mLTS method was further demonstrated with in vivo 23 Na-MRI data, showing significantly lower regression errors (ie root mean squared error) as compared with conventional LR methods (p < 0.001). The results of simulation and in vivo experiments revealed that 3D-mLTS is superior for determining under- or overestimated TSCs while preserving anatomical details. This suggests that the 3D-mLTS method is well suited for the accurate determination of TSC, especially in small focal lesions associated with pathological conditions.

Bone regeneration potential of allogeneic or autogeneic mesenchymal stem cells loaded onto cancellous bone granules in a rabbit radial defect model.

For developing a clinically effective bone regeneration strategy, we compare the bone regeneration potential of cultured allogeneic bone marrow-derived mesenchymal stem cells (BM-MSCs) and of autologous BM-MSCs loaded onto allogeneic cancellous bone granule scaffolds. A critical-sized segmental bone defect was made at the mid-shaft of both radiuses in 19 New Zealand White rabbits (NWRs). In the experimental group, allogeneic BM-MSCs loaded onto small-sized allogeneic cancellous bone granules (300~700 um in diameter) were implanted in one side of a bone defect. In the control group, autologous BM-MSCs loaded onto allogeneic cancellous granules were grafted in the other side. Bone regeneration was assessed by radiographic evaluation at 4, 8, 12 and 16 weeks post-implantation and by micro-computed tomography (micro-CT) and histological evaluation at 8 and 16 weeks. The experimental groups showed lower bone quantity indices (BQIs) than the control groups at 12 and 16 weeks (p < 0.05), although no significant difference was observed at 4 and 8 weeks (p > 0.05). Micro-CT analysis revealed that both groups had similar mean total bone volume and other parameters including trabecular thickness, number and separation at either 8 or 16 weeks. Only bone surface area revealed less area in the experimental group at 16 weeks. Histological evaluation of 8-week and 16-week specimens showed similar biologic processes of new bone formation and maturation. There was no inflammatory reaction indicating an adverse immune response in both allogeneic and autologous MSC groups. In conclusion, allogeneic BM-MSCs loaded onto allogeneic cancellous bone granules had comparable bone regeneration potential to autologous BM-MSCs in a rabbit radial defect model.

Quick and accurate measures in negative pressure pulmonary edema: a guideline for orthognathic surgeons.

Perioperative pulmonary edema is a rare complication of maxillofacial surgery. However, this potentially fatal complication may arise during any maxillofacial surgery. Negative pressure pulmonary edema can be caused by upper airway obstruction after operation. When this phenomenon arises, if not treated properly, it progresses rapidly causing fatal outcomes. Because orthognathic surgery is performed mostly on healthy and young patients, surgeons and anesthesiologists might neglect the possibility of such complications. Therefore, we must always take into consideration the possibility of negative pressure pulmonary edema. Careful observation of the patient; accurate knowledge; and active, quick, and noninvasive safe measures are required to treat this malady when it does occur. We report a case of negative pressure pulmonary edema after orthognathic surgery and its successful treatment.

Domain transformation learning for MR image reconstruction from dual domain input.

Medical images are acquired through diverse imaging systems, with each system employing specific image reconstruction techniques to transform sensor data into images. In MRI, sensor data (i.e., k-space data) is encoded in the frequency domain, and fully sampled k-space data is transformed into an image using the inverse Fourier Transform. However, in efforts to reduce acquisition time, k-space is often subsampled, necessitating a sophisticated image reconstruction method beyond a simple transform. The proposed approach addresses this challenge by training a model to learn domain transform, generating the final image directly from undersampled k-space input. Significantly, to improve the stability of reconstruction from randomly subsampled k-space data, folded images are incorporated as supplementary inputs in the dual-input ETER-net. Moreover, modifications are made to the formation of inputs for the bi-RNN stages to accommodate non-fixed k-space trajectories. Experimental validation, encompassing both regular and irregular sampling trajectories, validates the method's effectiveness. The results demonstrated superior performance, measured by PSNR, SSIM, and VIF, across acceleration factors of 4 and 8. In summary, the dual-input ETER-net emerges as an effective both regular and irregular sampling trajectories, and accommodating diverse acceleration factors.

Super-resolution track-density imaging of thalamic substructures: comparison with high-resolution anatomical magnetic resonance imaging at 7.0T.

The thalamus is one of the most important brain structures, with strong connections between subcortical and cortical areas of the brain. Most of the incoming information to the cortex passes through the thalamus. Accurate identification of substructures of the thalamus is therefore of great importance for the understanding of human brain connectivity. Direct visualization of thalamic substructures, however, is not easily achieved with currently available magnetic resonance imaging (MRI), including ultra-high field MRI such as 7.0T, mainly due to the limited contrast between the relevant structures. Recently, improvements in ultra-high field 7.0T MRI have opened the possibility of observing thalamic substructures by well-adjusted high-resolution T1 -weighted imaging. Moreover, the recently developed super-resolution track-density imaging (TDI) technique, based on results from whole-brain fiber-tracking, produces images with sub-millimeter resolution. These two methods enable us to show markedly improved anatomical detail of the substructures of the thalamus, including their detailed locations and directionality. In this study, we demonstrate the role of TDI for the visualization of the substructures of the thalamic nuclei, and relate these images to T1-weighted imaging at 7.0T MRI.

Distortion correction in EPI at ultra-high-field MRI using PSF mapping with optimal combination of shift detection dimension.

Despite its wide use, echo-planar imaging (EPI) suffers from geometric distortions due to off-resonance effects, i.e., strong magnetic field inhomogeneity and susceptibility. This article reports a novel method for correcting the distortions observed in EPI acquired at ultra-high-field such as 7 T. Point spread function (PSF) mapping methods have been proposed for correcting the distortions in EPI. The PSF shift map can be derived either along the nondistorted or the distorted coordinates. Along the nondistorted coordinates more information about compressed areas is present but it is prone to PSF-ghosting artifacts induced by large k-space shift in PSF encoding direction. In contrast, shift maps along the distorted coordinates contain more information in stretched areas and are more robust against PSF-ghosting. In ultra-high-field MRI, an EPI contains both compressed and stretched regions depending on the B0 field inhomogeneity and local susceptibility. In this study, we present a new geometric distortion correction scheme, which selectively applies the shift map with more information content. We propose a PSF-ghost elimination method to generate an artifact-free pixel shift map along nondistorted coordinates. The proposed method can correct the effects of the local magnetic field inhomogeneity induced by the susceptibility effects along with the PSF-ghost artifact cancellation. We have experimentally demonstrated the advantages of the proposed method in EPI data acquisitions in phantom and human brain using 7-T MRI.

Association of interferon gamma polymorphism with ossification of the posterior longitudinal ligament in the korean population.

In this study, we investigated whether genetic polymorphisms of the interferon gamma (IFNG) gene were associated with the susceptibility of ossification of the posterior longitudinal ligament (OPLL) in the Korean population. To observe the association between the IFNG gene and the susceptibility of OPLL, we genotyped 135 OPLL patients and 222 control subjects for a single nucleotide polymorphism (SNP, rs2430561) and a microsatellite (CA(n) repeats, rs3138557) located in the first intron of the IFNG gene, using the direct sequencing and gene scan method. The numbers of microsatellites (CA(13) and CA(15)) were significantly changed in the OPLL patients. A combined analysis of the genotype of rs2430561 and the number of microsatellites revealed that the OPLL was associated with frequencies of CA(13)-AA, CA(15)-AA and CA(15)-AT. Our results suggest that the IFNG gene may be one of the factors determining the OPLL in the Korean population. However, larger collaborative and biological studies are needed to confirm these results.

A correction method for streak artifacts in gradient-echo EPI using spin-echo EPI reference data.

OBJECTIVE: To analyze the streak artifacts in a gradient-echo echo planar imaging (GE-EPI) sequence and to propose a correction method for the Nyquist ghost artifacts that does not cause streak artifacts in the GE-EPI imaging. MATERIALS AND METHODS: Several GE-EPI imaging experiments with various reference scans, using both GE-EPI and SE-EPI scan data, were performed to analyze the streak artifacts and to investigate the spin dephasing phenomena of the GE-EPI reference scan. In addition, the analysis based on the spin dephasing was undertaken in order to demonstrate that the SE-EPI reference data can be used for the correction of the GE-EPI main scan data. RESULTS: The experimental results confirmed that the improvement of the reference data using either signal averaging or a large flip angle cannot guarantee perfect correction of the streak artifact if the noise is not completely removed. Due to the main field inhomogeneity, the spins of the GE-EPI reference data were dephased in multiple echo signals. The proposed correction method, which uses a SE-EPI reference scan for the GE-EPI images, eliminates the N/2 ghost artifacts without producing streak artifacts. CONCLUSION: It is believed that the proposed phase error correction scheme can improve the EPI performance in high field MRIs with higher magnetic field inhomogeneities.

Local delivery of alendronate eluting chitosan scaffold can effectively increase osteoblast functions and inhibit osteoclast differentiation.

The aim of this study was to investigate the effect of alendronate released from chitosan scaffolds on enhancement of osteoblast functions and inhibition of osteoclast differentiation in vitro. The surface and cell morphologies of chitosan scaffolds and alendronate-loaded chitosan scaffolds were characterized by variable pressure field emission scanning electron microscope (VP-FE-SEM). Alendronate was released in a sustained manner. For evaluating osteoblast functions in MG-63 cells, we investigated cell proliferation, alkaline phosphatase (ALP) activity, and calcium deposition. Furthermore, for evaluating inhibition of osteoclast differentiation in RAW 264.7 cells, we investigated tartrate-resistant acid phosphatase (TRAP) activity, TRAP staining, and gene expressions. The in vitro studies revealed that osteoblasts grown on alendronate-loaded chitosan scaffold showed a significant increment in cell proliferation, ALP activity, and calcium deposition as compared to those grown on chitosan scaffolds. In addition, the in vitro study showed that osteoclast differentiation in RAW 264.7 cells cultured on alendronate-loaded chitosan scaffolds was greatly inhibited as compared to those cultured on chitosan scaffolds by the results of TRAP activity, TRAP staining, and gene expressions. Taken together, alendronate-loaded chitosan scaffolds could achieve the dual functions of improvement in osteoblast functions and inhibition of osteoclast differentiation. Thus, alendronate-eluting chitosan substrates are promising materials for enhancing osteoblast functions and inhibiting osteoclast differentiation in orthopedic and dental fields.

Transtendon arthroscopic repair of high grade partial-thickness articular surface tears of the rotator cuff with biceps tendon augmentation: technical note and preliminary results.

INTRODUCTION: Partial articular surface of the rotator cuff tendon tears has been recognized as a source of treatable shoulder pain and a precursory pathology for full-thickness tendon tears. Arthroscopic rotator cuff repair is a possible surgical method of treatment. Recent data have shown that the treating partial-thickness rotator cuff repairs with transtendon technique shows good clinical outcome. The use of this technique enables the reconstitution of the tendon with complete reconstruction of its footprint without damaging its intact bursal part. In cases of high grade partial articular-sided degenerative rotator cuff tears (involving >50% of the tendon) in older patients, there is a possibility of poor healing or re-tear of the rotator cuff repair, which may be associated with poor tendon quality and substantial thinning of the rotator cuff, subsequently revision surgery in these patients will be demanding. METHODS: To mitigate these problems, we describe here a new arthroscopic transtendon repair technique with tenotomized long head biceps tendon augmentation for high grade partial articular rotator cuff tear with the goal of providing increase tendon healing, as well as to minimize the probability of failure of the construct and to improve the clinical outcomes. RESULTS: The clinical results of the first 39 consecutive patients are reported showing significant decrease in pain and improved shoulder scores, as well as the post-operative range of motion and with no cases of re-tear of the rotator cuff tendon.

Baseline Correction of the Human 1H MRS(I) Spectrum Using T2* Selective Differential Operators in the Frequency Domain.

The baseline distortion caused by water and fat signals is a crucial issue in the 1H MRS(I) study of the human brain. This paper suggests an effective and reliable preprocessing technique to calibrate the baseline distortion caused by the water and fat signals exhibited in the MRS spectral signal. For the preprocessing, we designed a T2* (or linewidth within the spectral signal) selective filter for the MRS(I) data based on differential filtering within the frequency domain. The number and types for the differential filtering were determined by comparing the T2* selectivity profile of each differential operator with the T2* profile of the metabolites to be suppressed within the MRS(I) data. In the performance evaluation of the proposed differential filtering, the simulation data for MRS spectral signals were used. Furthermore, the spectral signal of the human 1H MRSI data obtained by 2D free induction decay chemical shift imaging with a typical water suppression technique was also used in the performance evaluation. The absolute values of the average of the filtered dataset were quantitatively analyzed using the LCModel software. With the suggested T2* selective (not frequency selective) filtering technique, in the simulated MRS data, we removed the metabolites from the simulated MRS(I) spectral signal baseline distorted by the water and fat signal observed in the most frequency band. Moreover, in the obtained MRSI data, the quantitative analysis results for the metabolites of interest showed notable improvement in the uncertainty estimation accuracy, the CRLB (Cramer-Rao Lower Bound) levels.

Bilateral superficial cervical plexus block for parathyroidectomy during pregnancy: A case report.

BACKGROUND: Primary hyperparathyroidism (PHPT) is the most common cause of pregnancy-related hypercalcemia. PHPT can cause maternal and fetal complications in pregnant women. General anesthesia for non-obstetric surgery in pregnant women is associated with maternal hazards and concerns regarding long-term neonatal neurocognitive effects. Surgical removal of the lesion in mid-pregnancy is currently the primary treatment option for pregnant patients with PHPT. However, the blood calcium concentration at which surgery should be considered remains under discussion due to the risk of miscarriage. CASE SUMMARY: A 31-year-old nulliparous woman at 11 wk of gestation was admitted to our hospital for parathyroidectomy. The patient had a history of intrauterine fetal death with unknown etiology at 16 wk of gestation 1 year prior. Her blood test results showed that the serum calcium level was elevated to 12.9 mg/dL, and the parathyroid hormone level was elevated to 157 pg/mL. In a neck ultrasound, it revealed a 0.8 cm × 1.5 cm sized oval, hypoechoic mass in the upper posterior of the left thyroid gland, which was compatible with parathyroid adenoma. Superficial cervical plexus block (SCPB) for parathyroidectomy was performed. After surgery, the obstetrician checked the status of the fetus, and there were no abnormal signs. Since then her calcium level returned to normal values after one week of surgery and a healthy male neonate of 2910 g was delivered vaginally at 38 wk of gestation. CONCLUSION: Our case suggests that SCPB can be an anesthetic option for parathyroidectomy during the first trimester of pregnancy.

Competition-level code generation with AlphaCode.

Programming is a powerful and ubiquitous problem-solving tool. Systems that can assist programmers or even generate programs themselves could make programming more productive and accessible. Recent transformer-based neural network models show impressive code generation abilities yet still perform poorly on more complex tasks requiring problem-solving skills, such as competitive programming problems. Here, we introduce AlphaCode, a system for code generation that achieved an average ranking in the top 54.3% in simulated evaluations on recent programming competitions on the Codeforces platform. AlphaCode solves problems by generating millions of diverse programs using specially trained transformer-based networks and then filtering and clustering those programs to a maximum of just 10 submissions. This result marks the first time an artificial intelligence system has performed competitively in programming competitions.