Heads-up no-limit Texas Hold’em presents significant challenges in AI due to its imperfect information and sparse reward system, with feedback provided only at the end of the game. While previous research has achieved notable successes, many of these approaches are computationally intensive, limiting their practicality in broader applications. In this work, we introduce A2CHoldem, a deep reinforcement learning model based on the advantage actor-critic algorithm, designed to address these challenges. A2CHoldem integrates several innovative strategies, including an enhanced state representation with continuous chip handling, and a custom reward function with intermediate rewards and penalties. During training, we applied various acceleration techniques, among them the incorporation of supervised learning approaches, to enhance convergence speed and stability. Experimental results indicate that A2CHoldem achieves 132.5 mbb/h against Slumbot, 58.2 mbb/h against OpenDeepStack (our reproduction of DeepStack), and 26.8 mbb/h against OpenAlphaHoldem (our reproduction of AlphaHoldem). These results demonstrate that A2CHoldem significantly outperforms all three models in terms of performance. Additionally, A2CHoldem’s single decision time is less than 0.1 seconds, which is substantially faster than OpenDeepStack.

DouDiZhu, as an imperfect information game, poses a great challenge to existing technologies because of its imperfect information, large action space and complex situation. To address these challenges, we present SPDou, an AI system specifically designed for DouDiZhu. The strategy learning process is enhanced by combining empirical strategies with model generation strategies using an Improved Sampling Proximal Policy Optimization (ISPPO) algorithm. In order to solve the problem of imperfect information, we treat it as observable information and feed it into the critic network of ISPPO. In addition, Reward Shaping Based on Hand Scoring and Difference Steps for Clearing Hand Mechanism(RSHSDSM) is designed to adapt the value of our hands according to the opponent’s role adaptation, which provides a reasonable feedback mechanism in a sparse reward environment. In order to improve the information extraction ability and training efficiency of the network, a new method of game state encoding is proposed, including Card Hand Coding, Legal Combination Coding and Size Coding. Experiments show that the SPDou proposed in this paper achieved an average winning rate of 65% in competing with other agents in various situations. SPDou performs better than the benchmark algorithms in terms of decision-making speed and it is about 75% faster than other agents. Moreover, through the Elo rating comparison, SPDou obtained significantly higher Elo scores than other agents.

Blind-spot denoising (BSD) method is a powerful paradigm for zero-shot image denoising by training models to predict masked target pixels from their neighbors. However, they struggle with real-world noise exhibiting strong local correlations, where efforts to suppress noise correlation often weaken pixel-value dependencies, adversely affecting denoising performance. This paper presents a theoretical analysis quantifying the impact of replacing masked pixels with observations exhibiting weaker noise correlation but potentially reduced similarity, revealing a trade-off that impacts the statistical risk of the estimation. Guided by this insight, we propose a computational scheme that replaces masked pixels with distant ones of similar appearance and lower noise correlation. This strategy improves the prediction by balancing noise suppression and structural consistency. Experiments confirm the effectiveness of our method, outperforming existing zero-shot BSD methods.