Reinforcement learning has achieved remarkable success across diverse scenarios. However, learning optimal policies within partially observable games remains a formidable challenge. Crucial privileged information in states is often shrouded during gameplay, yet ideally, it should be accessible and exploitable during training. Previous studies have concentrated on formulating policies based wholly on partial observations or oracle states. Nevertheless, these approaches often face hindrances in attaining effective generalization. To surmount this challenge, we propose the actor–cross-critic (ACC) learning framework, integrating both partial observations and oracle states. ACC achieves this by coordinating two critics and invoking a maximization operation mechanism to switch between them dynamically. This approach encourages the selection of the higher values when computing advantages within the actor–critic framework, thereby accelerating learning and mitigating bias under partial observability. Some theoretical analyses show that ACC exhibits better learning ability toward optimal policies than actor–critic learning using the oracle states. We highlight its superior performance through comprehensive evaluations in decision-making tasks, such asQuestBall,Minigrid, andAtari, and the challenging card gameDouDizhu.

Partial observability in environments poses significant challenges that impede the formation of effective policies in reinforcement learning. Prior research has shown that borrowing the complete state information can enhance sample efficiency. This strategy, however, frequently encounters unstable learning with high variance in practical applications due to the over-reliance on complete information. This paper introduces DCRL, a Dual Critic Reinforcement Learning framework designed to adaptively harness full-state information during training to reduce variance for optimized online performance. In particular, DCRL incorporates two distinct critics: an oracle critic with access to complete state information and a standard critic functioning within the partially observable context. It innovates a synergistic strategy to meld the strengths of the oracle critic for efficiency improvement and the standard critic for variance reduction, featuring a novel mechanism for seamless transition and weighting between them. We theoretically prove that DCRL mitigates the learning variance while maintaining unbiasedness. Extensive experimental analyses across the Box2D and Box3D environments have verified DCRL's superior performance. The source code is available in the supplementary.

Owing to the unremitting efforts from a few institutes, researchers have recently made significant progress in designing superhuman artificial intelligence (AI) in no-limit Texas hold'em (NLTH), the primary testbed for large-scale imperfect-information game research. However, it remains challenging for new researchers to study this problem since there are no standard benchmarks for comparing with existing methods, which hinders further developments in this research area. This work presents OpenHoldem, an integrated benchmark for large-scale imperfect-information game research using NLTH. OpenHoldem makes three main contributions to this research direction: 1) a standardized evaluation protocol for thoroughly evaluating different NLTH AIs; 2) four publicly available strong baselines for NLTH AI; and 3) an online testing platform with easy-to-use APIs for public NLTH AI evaluation. We will publicly release OpenHoldem and hope it facilitates further studies on the unsolved theoretical and computational issues in this area and cultivates crucial research problems like opponent modeling and human-computer interactive learning.

Solving hard exploration tasks with sparse rewards is notoriously challenging in reinforcement learning (RL), which needs to address two key issues simultaneously: exploiting past successful experiences and exploring the unknown environment. Many prior works take expert demonstrations as successful experiences and learn to imitate them directly. However, these demonstrations are often not available in practice. Recently, curiosity-driven RL methods provide intrinsic rewards, encouraging the agent to explore states with high novelty. Nonetheless, they lack a mechanism for leveraging past good experiences effectively. This work presents a Pseudo Value Network Distillation (PVND) framework to balance the RL agent's exploitative and exploratory behaviors effectively and automatically. In particular, PVND learns to set high exploitation bonuses to the critical states in rewarded trajectories from past experiences and high exploration bonuses to the novel states that agents rarely visit during exploration. We theoretically demonstrate that PVND gives larger positive intrinsic rewards to more critical states. Furthermore, PVND automatically finds meaningful and critical hierarchical sub-tasks for agents to accomplish the final goal progressively. Competitive results in several hard exploration sparse reward problems have verified its effectiveness and efficiency.

Heads-up no-limit Texas hold’em (HUNL) is the quintessential game with imperfect information. Representative priorworks like DeepStack and Libratus heavily rely on counter-factual regret minimization (CFR) and its variants to tackleHUNL. However, the prohibitive computation cost of CFRiteration makes it difficult for subsequent researchers to learnthe CFR model in HUNL and apply it in other practical applications. In this work, we present AlphaHoldem, a high-performance and lightweight HUNL AI obtained with an end-to-end self-play reinforcement learning framework. The proposed framework adopts a pseudo-siamese architecture to directly learn from the input state information to the output actions by competing the learned model with its different historical versions. The main technical contributions include anovel state representation of card and betting information, amultitask self-play training loss function, and a new modelevaluation and selection metric to generate the final model.In a study involving 100,000 hands of poker, AlphaHoldemdefeats Slumbot and DeepStack using only one PC with threedays training. At the same time, AlphaHoldem only takes 2.9milliseconds for each decision-making using only a singleGPU, more than 1,000 times faster than DeepStack. We release the history data among among AlphaHoldem, Slumbot,and top human professionals in the author’s GitHub repository to facilitate further studies in this direction.

Despite significant breakthroughs in developing gaming artificial intelligence (AI), Mahjong remains quite challenging as a popular multi-player imperfect information game. Compared with games such as Go and Texas Hold’em, Mahjong has much more invisible information, unfixed game order, and a complicated scoring system, resulting in high randomness and variance of the rewarding signals during the reinforcement learning process. This paper presents a Mahjong AI by introducing Reward Variance Reduction (RVR) into a new self-play deep reinforcement learning algorithm. RVR handles the invisibility via a relative value network which leverages the global information to guide the model to converge to the optimal strategy under an oracle with perfect information. Moreover, RVR improves the training stability using an expected reward network to adapt to the complex, dynamic, and highly stochastic reward environment. Extensive experimental results show that RVR significantly reduces the variance in Mahjong AI training and improves the model performance. After only three days of self-play training on a single server with 8 GPUs, RVR defeats 62.5% opponents on the Botzone platform.

This paper studies the challenging problem of reinforcement learning (RL) in hard exploration tasks with sparse rewards. It focuses on the exploration stage before the agent gets the first positive reward, in which case, traditional RL algorithms with simple exploration strategies often work poorly. Unlike previous methods using some attribute of a single state as the intrinsic reward to encourage exploration, this work leverages the social influence between different states to permit more efficient exploration. It introduces a general intrinsic reward construction method to evaluate the social influence of states dynamically. Three kinds of social influence are introduced for a state: conformity, power, and authority. By measuring the state’s social influence, agents quickly find the focus state during the exploration process. The proposed RL framework with state social influence evaluation works well in hard exploration task. Extensive experimental analyses and comparisons in Grid Maze and many hard exploration Atari 2600 games demonstrate its high exploration efficiency.

This work considers the problem of deep reinforcement learning (RL) with long time dependencies and sparse rewards, as are found in many hard exploration games. A graph-based representation is proposed to allow an agent to perform self-navigation for environmental exploration. The graph representation not only effectively models the environment structure, but also efficiently traces the agent state changes and the corresponding actions. By encouraging the agent to earn a new influence-based curiosity reward for new game observations, the whole exploration task is divided into sub-tasks, which are effectively solved using a unified deep RL model. Experimental evaluations on hard exploration Atari Games demonstrate the effectiveness of the proposed method. The source code and learned models will be released to facilitate further studies on this problem.

Experience replay plays a crucial role in Reinforcement Learning (RL), enabling the agent to remember and reuse experience from the past. Most previous methods sample experience transitions using simple heuristics like uniformly sampling or prioritizing those good ones. Since humans can learn from both good and bad experiences, more sophisticated experience replay algorithms need to be developed. Inspired by the potential energy in physics, this work introduces the artificial potential field into experience replay and develops Potentialized Experience Replay (PotER) as a new and effective sampling algorithm for RL in hard exploration tasks with sparse rewards. PotER defines a potential energy function for each state in experience replay and helps the agent to learn from both good and bad experiences using intrinsic state supervision. PotER can be combined with different RL algorithms as well as the self-imitation learning algorithm. Experimental analyses and comparisons on multiple challenging hard exploration environments have verified its effectiveness and efficiency.