I‐Chen Wu

Li-Cheng Lan, Huan Zhang, Ti-Rong Wu et al.

The success of AlphaZero (AZ) has demonstrated that neural-network-based Go AIs can surpass human performance by a large margin. Given that the state space of Go is extremely large and a human player can play the game from any legal state, we ask whether adversarial states exist for Go AIs that may lead them to play surprisingly wrong actions. In this paper, we first extend the concept of adversarial examples to the game of Go: we generate perturbed states that are ``semantically'' equivalent to the original state by adding meaningless moves to the game, and an adversarial state is a perturbed state leading to an undoubtedly inferior action that is obvious even for Go beginners. However, searching the adversarial state is challenging due to the large, discrete, and non-differentiable search space. To tackle this challenge, we develop the first adversarial attack on Go AIs that can efficiently search for adversarial states by strategically reducing the search space. This method can also be extended to other board games such as NoGo. Experimentally, we show that the actions taken by both Policy-Value neural network (PV-NN) and Monte Carlo tree search (MCTS) can be misled by adding one or two meaningless stones; for example, on 58\% of the AlphaGo Zero self-play games, our method can make the widely used KataGo agent with 50 simulations of MCTS plays a losing action by adding two meaningless stones. We additionally evaluated the adversarial examples found by our algorithm with amateur human Go players and 90\% of examples indeed lead the Go agent to play an obviously inferior action. Our code is available at \url{https://PaperCode.cc/GoAttack}.

Ti-Rong Wu, Hung Guei, Ting Han Wei et al.

Game solving is a similar, yet more difficult task than mastering a game. Solving a game typically means to find the game-theoretic value (outcome given optimal play), and optionally a full strategy to follow in order to achieve that outcome. The AlphaZero algorithm has demonstrated super-human level play, and its powerful policy and value predictions have also served as heuristics in game solving. However, to solve a game and obtain a full strategy, a winning response must be found for all possible moves by the losing player. This includes very poor lines of play from the losing side, for which the AlphaZero self-play process will not encounter. AlphaZero-based heuristics can be highly inaccurate when evaluating these out-of-distribution positions, which occur throughout the entire search. To address this issue, this paper investigates applying online fine-tuning while searching and proposes two methods to learn tailor-designed heuristics for game solving. Our experiments show that using online fine-tuning can solve a series of challenging 7x7 Killall-Go problems, using only 23.54% of computation time compared to the baseline without online fine-tuning. Results suggest that the savings scale with problem size. Our method can further be extended to any tree search algorithm for problem solving. Our code is available at https://rlg.iis.sinica.edu.tw/papers/neurips2023-online-fine-tuning-solver.

Chung-Chin Shih, Ting Han Wei, Ti-Rong Wu et al.

This paper describes a Relevance-Zone pattern table (RZT) that can be used to replace a traditional transposition table. An RZT stores exact game values for patterns that are discovered during a Relevance-Zone-Based Search (RZS), which is the current state-of-the-art in solving L&D problems in Go. Positions that share the same pattern can reuse the same exact game value in the RZT. The pattern matching scheme for RZTs is implemented using a radix tree, taking into consideration patterns with different shapes. To improve the efficiency of table lookups, we designed a heuristic that prevents redundant lookups. The heuristic can safely skip previously queried patterns for a given position, reducing the overhead to 10% of the original cost. We also analyze the time complexity of the RZT both theoretically and empirically. Experiments show the overhead of traversing the radix tree in practice during lookup remain flat logarithmically in relation to the number of entries stored in the table. Experiments also show that the use of an RZT instead of a traditional transposition table significantly reduces the number of searched nodes on two data sets of 7x7 and 19x19 L&D Go problems.

Kuo-Hao Ho, Ping-Chun Hsieh, Chiu-Chou Lin et al.

In this paper, we propose a new approach called Adaptive Behavioral Costs in Reinforcement Learning (ABC-RL) for training a human-like agent with competitive strength. While deep reinforcement learning agents have recently achieved superhuman performance in various video games, some of these unconstrained agents may exhibit actions, such as shaking and spinning, that are not typically observed in human behavior, resulting in peculiar gameplay experiences. To behave like humans and retain similar performance, ABC-RL augments behavioral limitations as cost signals in reinforcement learning with dynamically adjusted weights. Unlike traditional constrained policy optimization, we propose a new formulation that minimizes the behavioral costs subject to a constraint of the value function. By leveraging the augmented Lagrangian, our approach is an approximation of the Lagrangian adjustment, which handles the trade-off between the performance and the human-like behavior. Through experiments conducted on 3D games in DMLab-30 and Unity ML-Agents Toolkit, we demonstrate that ABC-RL achieves the same performance level while significantly reducing instances of shaking and spinning. These findings underscore the effectiveness of our proposed approach in promoting more natural and human-like behavior during gameplay.

Hoang-Giang Cao, Weihao Zeng, I-Chen Wu

Picking cluttered general objects is a challenging task due to the complex geometries and various stacking configurations. Many prior works utilize pose estimation for picking, but pose estimation is difficult on cluttered objects. In this paper, we propose Cluttered Objects Descriptors (CODs), a dense cluttered objects descriptor that can represent rich object structures, and use the pre-trained CODs network along with its intermediate outputs to train a picking policy. Additionally, we train the policy with reinforcement learning, which enable the policy to learn picking without supervision. We conduct experiments to demonstrate that our CODs is able to consistently represent seen and unseen cluttered objects, which allowed for the picking policy to robustly pick cluttered general objects. The resulting policy can pick 96.69% of unseen objects in our experimental environment which is twice as cluttered as the training scenarios.

Kuo-Hao Ho, Ruei-Yu Jheng, Ji-Han Wu et al.

Job-shop scheduling problem (JSP) is a mathematical optimization problem widely used in industries like manufacturing, and flexible JSP (FJSP) is also a common variant. Since they are NP-hard, it is intractable to find the optimal solution for all cases within reasonable times. Thus, it becomes important to develop efficient heuristics to solve JSP/FJSP. A kind of method of solving scheduling problems is construction heuristics, which constructs scheduling solutions via heuristics. Recently, many methods for construction heuristics leverage deep reinforcement learning (DRL) with graph neural networks (GNN). In this paper, we propose a new approach, named residual scheduling, to solving JSP/FJSP. In this new approach, we remove irrelevant machines and jobs such as those finished, such that the states include the remaining (or relevant) machines and jobs only. Our experiments show that our approach reaches state-of-the-art (SOTA) among all known construction heuristics on most well-known open JSP and FJSP benchmarks. In addition, we also observe that even though our model is trained for scheduling problems of smaller sizes, our method still performs well for scheduling problems of large sizes. Interestingly in our experiments, our approach even reaches zero gap for 49 among 50 JSP instances whose job numbers are more than 150 on 20 machines.

Hoang-Giang Cao, Weihao Zeng, I-Chen Wu

It is crucial to address the following issues for ubiquitous robotics manipulation applications: (a) vision-based manipulation tasks require the robot to visually learn and understand the object with rich information like dense object descriptors; and (b) sim-to-real transfer in robotics aims to close the gap between simulated and real data. In this paper, we present Sim-to-Real Dense Object Nets (SRDONs), a dense object descriptor that not only understands the object via appropriate representation but also maps simulated and real data to a unified feature space with pixel consistency. We proposed an object-to-object matching method for image pairs from different scenes and different domains. This method helps reduce the effort of training data from real-world by taking advantage of public datasets, such as GraspNet. With sim-to-real object representation consistency, our SRDONs can serve as a building block for a variety of sim-to-real manipulation tasks. We demonstrate in experiments that pre-trained SRDONs significantly improve performances on unseen objects and unseen visual environments for various robotic tasks with zero real-world training.

Hung Guei

2048 is a single-player stochastic puzzle game. This intriguing and addictive game has been popular worldwide and has attracted researchers to develop game-playing programs. Due to its simplicity and complexity, 2048 has become an interesting and challenging platform for evaluating the effectiveness of machine learning methods. This dissertation conducts comprehensive research on reinforcement learning and computer game algorithms for 2048. First, this dissertation proposes optimistic temporal difference learning, which significantly improves the quality of learning by employing optimistic initialization to encourage exploration for 2048. Furthermore, based on this approach, a state-of-the-art program for 2048 is developed, which achieves the highest performance among all learning-based programs, namely an average score of 625377 points and a rate of 72% for reaching 32768-tiles. Second, this dissertation investigates several techniques related to 2048, including the n-tuple network ensemble learning, Monte Carlo tree search, and deep reinforcement learning. These techniques are promising for further improving the performance of the current state-of-the-art program. Finally, this dissertation discusses pedagogical applications related to 2048 by proposing course designs and summarizing the teaching experience. The proposed course designs adopt 2048-like games as materials for beginners to learn reinforcement learning and computer game algorithms. The courses have been successfully applied to graduate-level students and received well by student feedback.

Isaac Wu, Michael Maslowski

As large language models (LLMs) become integrated into various sensitive applications, prompt injection, the use of prompting to induce harmful behaviors from LLMs, poses an ever increasing risk. Prompt injection attacks can cause LLMs to leak sensitive data, spread misinformation, and exhibit harmful behaviors. To defend against these attacks, we propose CourtGuard, a locally-runnable, multiagent prompt injection classifier. In it, prompts are evaluated in a court-like multiagent LLM system, where a "defense attorney" model argues the prompt is benign, a "prosecution attorney" model argues the prompt is a prompt injection, and a "judge" model gives the final classification. CourtGuard has a lower false positive rate than the Direct Detector, an LLM as-a-judge. However, CourtGuard is generally a worse prompt injection detector. Nevertheless, this lower false positive rate highlights the importance of considering both adversarial and benign scenarios for the classification of a prompt. Additionally, the relative performance of CourtGuard in comparison to other prompt injection classifiers advances the use of multiagent systems as a defense against prompt injection attacks. The implementations of CourtGuard and the Direct Detector with full prompts for Gemma-3-12b-it, Llama-3.3-8B, and Phi-4-mini-instruct are available at https://github.com/isaacwu2000/CourtGuard.

Ti-Rong Wu, Ting-Han Wei, I-Chen Wu

AlphaZero has been very successful in many games. Unfortunately, it still consumes a huge amount of computing resources, the majority of which is spent in self-play. Hyperparameter tuning exacerbates the training cost since each hyperparameter configuration requires its own time to train one run, during which it will generate its own self-play records. As a result, multiple runs are usually needed for different hyperparameter configurations. This paper proposes using population based training (PBT) to help tune hyperparameters dynamically and improve strength during training time. Another significant advantage is that this method requires a single run only, while incurring a small additional time cost, since the time for generating self-play records remains unchanged though the time for optimization is increased following the AlphaZero training algorithm. In our experiments for 9x9 Go, the PBT method is able to achieve a higher win rate for 9x9 Go than the baselines, each with its own hyperparameter configuration and trained individually. For 19x19 Go, with PBT, we are able to obtain improvements in playing strength. Specifically, the PBT agent can obtain up to 74% win rate against ELF OpenGo, an open-source state-of-the-art AlphaZero program using a neural network of a comparable capacity. This is compared to a saturated non-PBT agent, which achieves a win rate of 47% against ELF OpenGo under the same circumstances.

Bo Yang Hsueh, Wei Li, I-Chen Wu

Learning rate scheduler has been a critical issue in the deep neural network training. Several schedulers and methods have been proposed, including step decay scheduler, adaptive method, cosine scheduler and cyclical scheduler. This paper proposes a new scheduling method, named hyperbolic-tangent decay (HTD). We run experiments on several benchmarks such as: ResNet, Wide ResNet and DenseNet for CIFAR-10 and CIFAR-100 datasets, LSTM for PAMAP2 dataset, ResNet on ImageNet and Fashion-MNIST datasets. In our experiments, HTD outperforms step decay and cosine scheduler in nearly all cases, while requiring less hyperparameters than step decay, and more flexible than cosine scheduler. Code is available at https://github.com/BIGBALLON/HTD.

Chiu-Chou Lin, Yu-Wei Shih, Kuei-Ting Kuo et al.

How can balance be quantified in game settings? This question is crucial for game designers, especially in player-versus-player (PvP) games, where analyzing the strength relations among predefined team compositions-such as hero combinations in multiplayer online battle arena (MOBA) games or decks in card games-is essential for enhancing gameplay and achieving balance. We have developed two advanced measures that extend beyond the simplistic win rate to quantify balance in zero-sum competitive scenarios. These measures are derived from win value estimations, which employ strength rating approximations via the Bradley-Terry model and counter relationship approximations via vector quantization, significantly reducing the computational complexity associated with traditional win value estimations. Throughout the learning process of these models, we identify useful categories of compositions and pinpoint their counter relationships, aligning with the experiences of human players without requiring specific game knowledge. Our methodology hinges on a simple technique to enhance codebook utilization in discrete representation with a deterministic vector quantization process for an extremely small state space. Our framework has been validated in popular online games, including Age of Empires II, Hearthstone, Brawl Stars, and League of Legends. The accuracy of the observed strength relations in these games is comparable to traditional pairwise win value predictions, while also offering a more manageable complexity for analysis. Ultimately, our findings contribute to a deeper understanding of PvP game dynamics and present a methodology that significantly improves game balance evaluation and design.

Nai-Chieh Huang, Ping-Chun Hsieh, Kuo-Hao Ho et al.

Proximal Policy Optimization algorithm employing a clipped surrogate objective (PPO-Clip) is a prominent exemplar of the policy optimization methods. However, despite its remarkable empirical success, PPO-Clip lacks theoretical substantiation to date. In this paper, we contribute to the field by establishing the first global convergence results of a PPO-Clip variant in both tabular and neural function approximation settings. Our findings highlight the $O(1/\sqrt{T})$ min-iterate convergence rate specifically in the context of neural function approximation. We tackle the inherent challenges in analyzing PPO-Clip through three central concepts: (i) We introduce a generalized version of the PPO-Clip objective, illuminated by its connection with the hinge loss. (ii) Employing entropic mirror descent, we establish asymptotic convergence for tabular PPO-Clip with direct policy parameterization. (iii) Inspired by the tabular analysis, we streamline convergence analysis by introducing a two-step policy improvement approach. This decouples policy search from complex neural policy parameterization using a regression-based update scheme. Furthermore, we gain deeper insights into the efficacy of PPO-Clip by interpreting these generalized objectives. Our theoretical findings also mark the first characterization of the influence of the clipping mechanism on PPO-Clip convergence. Importantly, the clipping range affects only the pre-constant of the convergence rate.

Wei-Chen Liao, Ti-Rong Wu, I-Chen Wu

Multi-agent reinforcement Learning (MARL) is often challenged by the sight range dilemma, where agents either receive insufficient or excessive information from their environment. In this paper, we propose a novel method, called Dynamic Sight Range Selection (DSR), to address this issue. DSR utilizes an Upper Confidence Bound (UCB) algorithm and dynamically adjusts the sight range during training. Experiment results show several advantages of using DSR. First, we demonstrate using DSR achieves better performance in three common MARL environments, including Level-Based Foraging (LBF), Multi-Robot Warehouse (RWARE), and StarCraft Multi-Agent Challenge (SMAC). Second, our results show that DSR consistently improves performance across multiple MARL algorithms, including QMIX and MAPPO. Third, DSR offers suitable sight ranges for different training steps, thereby accelerating the training process. Finally, DSR provides additional interpretability by indicating the optimal sight range used during training. Unlike existing methods that rely on global information or communication mechanisms, our approach operates solely based on the individual sight ranges of agents. This approach offers a practical and efficient solution to the sight range dilemma, making it broadly applicable to real-world complex environments.

Chiu-Chou Lin, I-Chen Wu

In competitive games, strength ratings like Elo are widely used to quantify player skill and support matchmaking by accounting for skill disparities better than simple win rate statistics. However, scalar ratings cannot handle complex intransitive relationships, such as counter strategies seen in Rock-Paper-Scissors. To address this, recent work introduced Neural Rating Table and Neural Counter Table, which combine scalar ratings with discrete counter categories to model intransitivity. While effective, these methods rely on neural network training and cannot perform real-time updates. In this paper, we propose an online update algorithm that extends Elo principles to incorporate real-time learning of counter categories. Our method dynamically adjusts both ratings and counter relationships after each match, preserving the explainability of scalar ratings while addressing intransitivity. Experiments on zero-sum competitive games demonstrate its practicality, particularly in scenarios without complex team compositions.

Yun-Jui Tsai, Ting Han Wei, Chi-Huang Lin et al.

Game solving is the process of finding the theoretical outcome for a game, assuming that all player choices are optimal. This paper focuses on a technique that can reduce the heuristic search space significantly for 7x7 Killall-Go. In Go and Killall-Go, live patterns are stones that are protected from opponent capture. Mutual life, also referred to as seki, is when both players' stones achieve life by sharing liberties with their opponent. Whichever player attempts to capture the opponent first will leave their own stones vulnerable. Therefore, it is critical to recognize seki patterns to avoid putting oneself in jeopardy. Recognizing seki can reduce the search depth significantly. In this paper, we enumerate all seki patterns up to a predetermined area size, then store these patterns into a seki table. This allows us to recognize seki during search, which significantly improves solving efficiency for the game of Killall-Go. Experiments show that a day-long, unsolvable position can be solved in 482 seconds with the addition of a seki table. For general positions, a 10% to 20% improvement in wall clock time and node count is observed.

Nhat-Minh Huynh, Hoang-Giang Cao, I-Chen Wu

Pommerman is a multi-agent environment that has received considerable attention from researchers in recent years. This environment is an ideal benchmark for multi-agent training, providing a battleground for two teams with communication capabilities among allied agents. Pommerman presents significant challenges for model-free reinforcement learning due to delayed action effects, sparse rewards, and false positives, where opponent players can lose due to their own mistakes. This study introduces a system designed to train multi-agent systems to play Pommerman using a combination of curriculum learning and population-based self-play. We also tackle two challenging problems when deploying the multi-agent training system for competitive games: sparse reward and suitable matchmaking mechanism. Specifically, we propose an adaptive annealing factor based on agents' performance to adjust the dense exploration reward during training dynamically. Additionally, we implement a matchmaking mechanism utilizing the Elo rating system to pair agents effectively. Our experimental results demonstrate that our trained agent can outperform top learning agents without requiring communication among allied agents.

Chung-Chin Shih, Ti-Rong Wu, Ting Han Wei et al.

Goal-achieving problems are puzzles that set up a specific situation with a clear objective. An example that is well-studied is the category of life-and-death (L&D) problems for Go, which helps players hone their skill of identifying region safety. Many previous methods like lambda search try null moves first, then derive so-called relevance zones (RZs), outside of which the opponent does not need to search. This paper first proposes a novel RZ-based approach, called the RZ-Based Search (RZS), to solving L&D problems for Go. RZS tries moves before determining whether they are null moves post-hoc. This means we do not need to rely on null move heuristics, resulting in a more elegant algorithm, so that it can also be seamlessly incorporated into AlphaZero's super-human level play in our solver. To repurpose AlphaZero for solving, we also propose a new training method called Faster to Life (FTL), which modifies AlphaZero to entice it to win more quickly. We use RZS and FTL to solve L&D problems on Go, namely solving 68 among 106 problems from a professional L&D book while a previous program solves 11 only. Finally, we discuss that the approach is generic in the sense that RZS is applicable to solving many other goal-achieving problems for board games.

Hung Guei, Lung-Pin Chen, I-Chen Wu

Temporal difference (TD) learning and its variants, such as multistage TD (MS-TD) learning and temporal coherence (TC) learning, have been successfully applied to 2048. These methods rely on the stochasticity of the environment of 2048 for exploration. In this paper, we propose to employ optimistic initialization (OI) to encourage exploration for 2048, and empirically show that the learning quality is significantly improved. This approach optimistically initializes the feature weights to very large values. Since weights tend to be reduced once the states are visited, agents tend to explore those states which are unvisited or visited few times. Our experiments show that both TD and TC learning with OI significantly improve the performance. As a result, the network size required to achieve the same performance is significantly reduced. With additional tunings such as expectimax search, multistage learning, and tile-downgrading technique, our design achieves the state-of-the-art performance, namely an average score of 625 377 and a rate of 72% reaching 32768 tiles. In addition, for sufficiently large tests, 65536 tiles are reached at a rate of 0.02%.

Nai-Chieh Huang, Ping-Chun Hsieh, Kuo-Hao Ho et al.

Policy optimization is a fundamental principle for designing reinforcement learning algorithms, and one example is the proximal policy optimization algorithm with a clipped surrogate objective (PPO-Clip), which has been popularly used in deep reinforcement learning due to its simplicity and effectiveness. Despite its superior empirical performance, PPO-Clip has not been justified via theoretical proof up to date. In this paper, we establish the first global convergence rate of PPO-Clip under neural function approximation. We identify the fundamental challenges of analyzing PPO-Clip and address them with the two core ideas: (i) We reinterpret PPO-Clip from the perspective of hinge loss, which connects policy improvement with solving a large-margin classification problem with hinge loss and offers a generalized version of the PPO-Clip objective. (ii) Based on the above viewpoint, we propose a two-step policy improvement scheme, which facilitates the convergence analysis by decoupling policy search from the complex neural policy parameterization with the help of entropic mirror descent and a regression-based policy update scheme. Moreover, our theoretical results provide the first characterization of the effect of the clipping mechanism on the convergence of PPO-Clip. Through experiments, we empirically validate the reinterpretation of PPO-Clip and the generalized objective with various classifiers on various RL benchmark tasks.

Chiu-Chou Lin, Wei-Chen Chiu, I-Chen Wu

On playing video games, different players usually have their own playstyles. Recently, there have been great improvements for the video game AIs on the playing strength. However, past researches for analyzing the behaviors of players still used heuristic rules or the behavior features with the game-environment support, thus being exhausted for the developers to define the features of discriminating various playstyles. In this paper, we propose the first metric for video game playstyles directly from the game observations and actions, without any prior specification on the playstyle in the target game. Our proposed method is built upon a novel scheme of learning discrete representations that can map game observations into latent discrete states, such that playstyles can be exhibited from these discrete states. Namely, we measure the playstyle distance based on game observations aligned to the same states. We demonstrate high playstyle accuracy of our metric in experiments on some video game platforms, including TORCS, RGSK, and seven Atari games, and for different agents including rule-based AI bots, learning-based AI bots, and human players.

Li-Cheng Lan, Meng-Yu Tsai, Ti-Rong Wu et al.

Monte Carlo tree search (MCTS) has achieved state-of-the-art results in many domains such as Go and Atari games when combining with deep neural networks (DNNs). When more simulations are executed, MCTS can achieve higher performance but also requires enormous amounts of CPU and GPU resources. However, not all states require a long searching time to identify the best action that the agent can find. For example, in 19x19 Go and NoGo, we found that for more than half of the states, the best action predicted by DNN remains unchanged even after searching 2 minutes. This implies that a significant amount of resources can be saved if we are able to stop the searching earlier when we are confident with the current searching result. In this paper, we propose to achieve this goal by predicting the uncertainty of the current searching status and use the result to decide whether we should stop searching. With our algorithm, called Dynamic Simulation MCTS (DS-MCTS), we can speed up a NoGo agent trained by AlphaZero 2.5 times faster while maintaining a similar winning rate. Also, under the same average simulation count, our method can achieve a 61% winning rate against the original program.

Yeong-Jia Roger Chu, Ting-Han Wei, Jin-Bo Huang et al.

Sim-to-real, a term that describes where a model is trained in a simulator then transferred to the real world, is a technique that enables faster deep reinforcement learning (DRL) training. However, differences between the simulator and the real world often cause the model to perform poorly in the real world. Domain randomization is a way to bridge the sim-to-real gap by exposing the model to a wide range of scenarios so that it can generalize to real-world situations. However, following domain randomization to train an autonomous car racing model with DRL can lead to undesirable outcomes. Namely, a model trained with randomization tends to run slower; a higher completion rate on the testing track comes at the expense of longer lap times. This paper aims to boost the robustness of a trained race car model without compromising racing lap times. For a training track and a testing track having the same shape (and same optimal paths), but with different lighting, background, etc., we first train a model (teacher model) that overfits the training track, moving along a near optimal path. We then use this model to teach a student model the correct actions along with randomization. With our method, a model with 18.4\% completion rate on the testing track is able to help teach a student model with 52\% completion. Moreover, over an average of 50 trials, the student is able to finish a lap 0.23 seconds faster than the teacher. This 0.23 second gap is significant in tight races, with lap times of about 10 to 12 seconds.

Li-Cheng Lan, Wei Li, Ting-Han Wei et al.

Many of the strongest game playing programs use a combination of Monte Carlo tree search (MCTS) and deep neural networks (DNN), where the DNNs are used as policy or value evaluators. Given a limited budget, such as online playing or during the self-play phase of AlphaZero (AZ) training, a balance needs to be reached between accurate state estimation and more MCTS simulations, both of which are critical for a strong game playing agent. Typically, larger DNNs are better at generalization and accurate evaluation, while smaller DNNs are less costly, and therefore can lead to more MCTS simulations and bigger search trees with the same budget. This paper introduces a new method called the multiple policy value MCTS (MPV-MCTS), which combines multiple policy value neural networks (PV-NNs) of various sizes to retain advantages of each network, where two PV-NNs f_S and f_L are used in this paper. We show through experiments on the game NoGo that a combined f_S and f_L MPV-MCTS outperforms single PV-NN with policy value MCTS, called PV-MCTS. Additionally, MPV-MCTS also outperforms PV-MCTS for AZ training.

Kai-Chun Hu, Chen-Huan Pi, Ting Han Wei et al.

In this paper, we point out a fundamental property of the objective in reinforcement learning, with which we can reformulate the policy gradient objective into a perceptron-like loss function, removing the need to distinguish between on and off policy training. Namely, we posit that it is sufficient to only update a policy $π$ for cases that satisfy the condition $A(\fracπμ-1)\leq0$, where $A$ is the advantage, and $μ$ is another policy. Furthermore, we show via theoretic derivation that a perceptron-like loss function matches the clipped surrogate objective for PPO. With our new formulation, the policies $π$ and $μ$ can be arbitrarily apart in theory, effectively enabling off-policy training. To examine our derivations, we can combine the on-policy PPO clipped surrogate (which we show to be equivalent with one instance of the new reformation) with the off-policy IMPALA method. We first verify the combined method on the OpenAI Gym pendulum toy problem. Next, we use our method to train a quadrotor position controller in a simulator. Our trained policy is efficient and lightweight enough to perform in a low cost micro-controller at a minimum update rate of 500 Hz. For the quadrotor, we show two experiments to verify our method and demonstrate performance: 1) hovering at a fixed position, and 2) tracking along a specific trajectory. In preliminary trials, we are also able to apply the method to a real-world quadrotor.

Wen-Jie Tseng, Jr-Chang Chen, I-Chen Wu et al.

This paper describes the application of comparison training (CT) for automatic feature weight tuning, with the final objective of improving the evaluation functions used in Chinese chess programs. First, we propose an n-tuple network to extract features, since n-tuple networks require very little expert knowledge through its large numbers of features, while simulta-neously allowing easy access. Second, we propose a novel evalua-tion method that incorporates tapered eval into CT. Experiments show that with the same features and the same Chinese chess program, the automatically tuned comparison training feature weights achieved a win rate of 86.58% against the weights that were hand-tuned. The above trained version was then improved by adding additional features, most importantly n-tuple features. This improved version achieved a win rate of 81.65% against the trained version without additional features.

Ti-Rong Wu, I-Chen Wu, Guan-Wun Chen et al.

This paper proposes a new approach to a novel value network architecture for the game Go, called a multi-labelled (ML) value network. In the ML value network, different values (win rates) are trained simultaneously for different settings of komi, a compensation given to balance the initiative of playing first. The ML value network has three advantages, (a) it outputs values for different komi, (b) it supports dynamic komi, and (c) it lowers the mean squared error (MSE). This paper also proposes a new dynamic komi method to improve game-playing strength. This paper also performs experiments to demonstrate the merits of the architecture. First, the MSE of the ML value network is generally lower than the value network alone. Second, the program based on the ML value network wins by a rate of 67.6% against the program based on the value network alone. Third, the program with the proposed dynamic komi method significantly improves the playing strength over the baseline that does not use dynamic komi, especially for handicap games. To our knowledge, up to date, no handicap games have been played openly by programs using value networks. This paper provides these programs with a useful approach to playing handicap games.

Kun-Hao Yeh, I-Chen Wu, Chu-Hsuan Hsueh et al.

Szubert and Jaskowski successfully used temporal difference (TD) learning together with n-tuple networks for playing the game 2048. However, we observed a phenomenon that the programs based on TD learning still hardly reach large tiles. In this paper, we propose multi-stage TD (MS-TD) learning, a kind of hierarchical reinforcement learning method, to effectively improve the performance for the rates of reaching large tiles, which are good metrics to analyze the strength of 2048 programs. Our experiments showed significant improvements over the one without using MS-TD learning. Namely, using 3-ply expectimax search, the program with MS-TD learning reached 32768-tiles with a rate of 18.31%, while the one with TD learning did not reach any. After further tuned, our 2048 program reached 32768-tiles with a rate of 31.75% in 10,000 games, and one among these games even reached a 65536-tile, which is the first ever reaching a 65536-tile to our knowledge. In addition, MS-TD learning method can be easily applied to other 2048-like games, such as Threes. Based on MS-TD learning, our experiments for Threes also demonstrated similar performance improvement, where the program with MS-TD learning reached 6144-tiles with a rate of 7.83%, while the one with TD learning only reached 0.45%.