Kenshi Abe

Kenshi Abe, Mitsuki Sakamoto, Atsushi Iwasaki

In this study, we consider a variant of the Follow the Regularized Leader (FTRL) dynamics in two-player zero-sum games. FTRL is guaranteed to converge to a Nash equilibrium when time-averaging the strategies, while a lot of variants suffer from the issue of limit cycling behavior, i.e., lack the last-iterate convergence guarantee. To this end, we propose mutant FTRL (M-FTRL), an algorithm that introduces mutation for the perturbation of action probabilities. We then investigate the continuous-time dynamics of M-FTRL and provide the strong convergence guarantees toward stationary points that approximate Nash equilibria under full-information feedback. Furthermore, our simulation demonstrates that M-FTRL can enjoy faster convergence rates than FTRL and optimistic FTRL under full-information feedback and surprisingly exhibits clear convergence under bandit feedback.

Kenshi Abe, Kaito Ariu, Mitsuki Sakamoto et al.

This paper proposes Mutation-Driven Multiplicative Weights Update (M2WU) for learning an equilibrium in two-player zero-sum normal-form games and proves that it exhibits the last-iterate convergence property in both full and noisy feedback settings. In the former, players observe their exact gradient vectors of the utility functions. In the latter, they only observe the noisy gradient vectors. Even the celebrated Multiplicative Weights Update (MWU) and Optimistic MWU (OMWU) algorithms may not converge to a Nash equilibrium with noisy feedback. On the contrary, M2WU exhibits the last-iterate convergence to a stationary point near a Nash equilibrium in both feedback settings. We then prove that it converges to an exact Nash equilibrium by iteratively adapting the mutation term. We empirically confirm that M2WU outperforms MWU and OMWU in exploitability and convergence rates.

Hakuei Yamada, Junpei Komiyama, Kenshi Abe et al.

This work addresses learning online fair division under uncertainty, where a central planner sequentially allocates items without precise knowledge of agents' values or utilities. Departing from conventional online algorithm, the planner here relies on noisy, estimated values obtained after allocating items. We introduce wrapper algorithms utilizing \textit{dual averaging}, enabling gradual learning of both the type distribution of arriving items and agents' values through bandit feedback. This approach enables the algorithms to asymptotically achieve optimal Nash social welfare in linear Fisher markets with agents having additive utilities. We establish regret bounds in Nash social welfare and empirically validate the superior performance of our proposed algorithms across synthetic and empirical datasets.

Yuu Jinnai, Tetsuro Morimura, Ukyo Honda et al.

Minimum Bayes Risk (MBR) decoding has been shown to be a powerful alternative to beam search decoding in a variety of text generation tasks. MBR decoding selects a hypothesis from a pool of hypotheses that has the least expected risk under a probability model according to a given utility function. Since it is impractical to compute the expected risk exactly over all possible hypotheses, two approximations are commonly used in MBR. First, it integrates over a sampled set of hypotheses rather than over all possible hypotheses. Second, it estimates the probability of each hypothesis using a Monte Carlo estimator. While the first approximation is necessary to make it computationally feasible, the second is not essential since we typically have access to the model probability at inference time. We propose Model-Based MBR (MBMBR), a variant of MBR that uses the model probability itself as the estimate of the probability distribution instead of the Monte Carlo estimate. We show analytically and empirically that the model-based estimate is more promising than the Monte Carlo estimate in text generation tasks. Our experiments show that MBMBR outperforms MBR in several text generation tasks, both with encoder-decoder models and with large language models.

Tetsuro Morimura, Mitsuki Sakamoto, Yuu Jinnai et al.

Reinforcement learning from human feedback (RLHF) plays a crucial role in aligning language models with human preferences. While the significance of dataset quality is generally recognized, explicit investigations into its impact within the RLHF framework, to our knowledge, have been limited. This paper addresses the issue of text quality within the preference dataset by focusing on direct preference optimization (DPO), an increasingly adopted reward-model-free RLHF method. We confirm that text quality significantly influences the performance of models optimized with DPO more than those optimized with reward-model-based RLHF. Building on this new insight, we propose an extension of DPO, termed filtered direct preference optimization (fDPO). fDPO uses a trained reward model to monitor the quality of texts within the preference dataset during DPO training. Samples of lower quality are discarded based on comparisons with texts generated by the model being optimized, resulting in a more accurate dataset. Experimental results demonstrate that fDPO enhances the final model performance. Our code is available at https://github.com/CyberAgentAILab/filtered-dpo.

Sho Shimoyama, Tetsuro Morimura, Kenshi Abe et al.

Dialog policies, which determine a system's action based on the current state at each dialog turn, are crucial to the success of the dialog. In recent years, reinforcement learning (RL) has emerged as a promising option for dialog policy learning (DPL). In RL-based DPL, dialog policies are updated according to rewards. The manual construction of fine-grained rewards, such as state-action-based ones, to effectively guide the dialog policy is challenging in multi-domain task-oriented dialog scenarios with numerous state-action pair combinations. One way to estimate rewards from collected data is to train the reward estimator and dialog policy simultaneously using adversarial learning (AL). Although this method has demonstrated superior performance experimentally, it is fraught with the inherent problems of AL, such as mode collapse. This paper first identifies the role of AL in DPL through detailed analyses of the objective functions of dialog policy and reward estimator. Next, based on these analyses, we propose a method that eliminates AL from reward estimation and DPL while retaining its advantages. We evaluate our method using MultiWOZ, a multi-domain task-oriented dialog corpus.

Yuu Jinnai, Tetsuro Morimura, Kaito Ariu et al.

Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) to human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking when the accuracy of the reward model is not high enough due to the quality or the quantity of the preference dataset. Because the reward model is an imperfect proxy for the true objective, over-optimizing its value can compromise its performance on the true objective. In this research, we propose MBR-BoN, a variant of BoN that aims to mitigate reward hacking at inference time by incorporating the Minimum Bayes Risk (MBR) objective as a proximity regularization term. We show empirically and analytically that the MBR objective quantifies the proximity of the response to the reference policy, serving as a proximity regularizer. We evaluate MBR-BoN on the AlpacaFarm and Anthropic's hh-rlhf datasets and show that it outperforms both BoN sampling and MBR decoding. We also evaluate MBR-BoN to generate a pairwise preference learning dataset for Direct Preference Optimization (DPO). Empirical results show that models trained on a dataset generated with MBR-BoN outperform those with vanilla BoN. Our code is available at https://github.com/CyberAgentAILab/regularized-bon

Tetsuro Morimura, Kazuhiro Ota, Kenshi Abe et al.

Policy gradient (PG) is a reinforcement learning (RL) approach that optimizes a parameterized policy model for an expected return using gradient ascent. While PG can work well even in non-Markovian environments, it may encounter plateaus or peakiness issues. As another successful RL approach, algorithms based on Monte Carlo Tree Search (MCTS), which include AlphaZero, have obtained groundbreaking results, especially in the game-playing domain. They are also effective when applied to non-Markov decision processes. However, the standard MCTS is a method for decision-time planning, which differs from the online RL setting. In this work, we first introduce Monte Carlo Tree Learning (MCTL), an adaptation of MCTS for online RL setups. We then explore a combined policy approach of PG and MCTL to leverage their strengths. We derive conditions for asymptotic convergence with the results of a two-timescale stochastic approximation and propose an algorithm that satisfies these conditions and converges to a reasonable solution. Our numerical experiments validate the effectiveness of the proposed methods.

Yuma Fujimoto, Kenshi Abe, Kaito Ariu

Feedback delays are inevitable in real-world multi-agent learning. They are known to severely degrade performance, and the convergence rate under delayed feedback is still unclear, even for bilinear games. This paper derives the rate of linear convergence of Weighted Optimistic Gradient Descent-Ascent (WOGDA), which predicts future rewards with extra optimism, in unconstrained bilinear games. To analyze the algorithm, we interpret it as an approximation of the Extra Proximal Point (EPP), which is updated based on farther future rewards than the classical Proximal Point (PP). Our theorems show that standard optimism (predicting the next-step reward) achieves linear convergence to the equilibrium at a rate $\exp(-Θ(t/m^{5}))$ after $t$ iterations for delay $m$. Moreover, employing extra optimism (predicting farther future reward) tolerates a larger step size and significantly accelerates the rate to $\exp(-Θ(t/(m^{2}\log m)))$. Our experiments also show accelerated convergence driven by the extra optimism and are qualitatively consistent with our theorems. In summary, this paper validates that extra optimism is a promising countermeasure against performance degradation caused by feedback delays.

Tsunehiko Tanaka, Kenshi Abe, Kaito Ariu et al.

Traditional approaches in offline reinforcement learning aim to learn the optimal policy that maximizes the cumulative reward, also known as return. It is increasingly important to adjust the performance of AI agents to meet human requirements, for example, in applications like video games and education tools. Decision Transformer (DT) optimizes a policy that generates actions conditioned on the target return through supervised learning and includes a mechanism to control the agent's performance using the target return. However, the action generation is hardly influenced by the target return because DT's self-attention allocates scarce attention scores to the return tokens. In this paper, we propose Return-Aligned Decision Transformer (RADT), designed to more effectively align the actual return with the target return. RADT leverages features extracted by paying attention solely to the return, enabling action generation to consistently depend on the target return. Extensive experiments show that RADT significantly reduces the discrepancies between the actual return and the target return compared to DT-based methods. Our code is available at https://github.com/CyberAgentAILab/radt

Yuki Ichihara, Yuu Jinnai, Tetsuro Morimura et al.

Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) with human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking. Since the reward model is an imperfect proxy for the true objective, an excessive focus on optimizing its value can lead to a compromise of its performance on the true objective. Previous work proposes Regularized BoN sampling (RBoN), a BoN sampling with regularization to the objective, and shows that it outperforms BoN sampling so that it mitigates reward hacking and empirically (Jinnai et al., 2024). However, Jinnai et al. (2024) introduce RBoN based on a heuristic and they lack the analysis of why such regularization strategy improves the performance of BoN sampling. The aim of this study is to analyze the effect of BoN sampling on regularization strategies. Using the regularization strategies corresponds to robust optimization, which maximizes the worst case over a set of possible perturbations in the proxy reward. Although the theoretical guarantees are not directly applicable to RBoN, RBoN corresponds to a practical implementation. This paper proposes an extension of the RBoN framework, called Stochastic RBoN sampling (SRBoN), which is a theoretically guaranteed approach to worst-case RBoN in proxy reward. We then perform an empirical evaluation using the AlpacaFarm and Anthropic's hh-rlhf datasets to evaluate which factors of the regularization strategies contribute to the improvement of the true proxy reward. In addition, we also propose another simple RBoN method, the Sentence Length Regularized BoN, which has a better performance in the experiment as compared to the previous methods.

Wataru Masaka, Mitsuki Sakamoto, Kenshi Abe et al.

We investigate how perturbation does and does not improve the Follow-the-Regularized-Leader (FTRL) algorithm in solving imperfect-information extensive-form games under sampling, where payoffs are estimated from sampled trajectories. While optimistic algorithms are effective under full feedback, they often become unstable in the presence of sampling noise. Payoff perturbation offers a promising alternative for stabilizing learning and achieving \textit{last-iterate convergence}. We present a unified framework for \textit{Perturbed FTRL} algorithms and study two variants: PFTRL-KL (standard KL divergence) and PFTRL-RKL (Reverse KL divergence), the latter featuring an estimator with both unbiasedness and conditional zero variance. While PFTRL-KL generally achieves equivalent or better performance across benchmark games, PFTRL-RKL consistently outperforms it in Leduc poker, whose structure is more asymmetric than the other games in a sense. Given the modest advantage of PFTRL-RKL, we design the second experiment to isolate the effect of conditional zero variance, showing that the variance-reduction property of RKL improve last-iterate performance.

Yuma Fujimoto, Kenshi Abe, Kaito Ariu

This study raises and addresses the problem of time-delayed feedback in learning in games. Because learning in games assumes that multiple agents independently learn their strategies, a discrepancy in optimization often emerges among the agents. To overcome this discrepancy, the prediction of the future reward is incorporated into algorithms, typically known as Optimistic Follow-the-Regularized-Leader (OFTRL). However, the time delay in observing the past rewards hinders the prediction. Indeed, this study firstly proves that even a single-step delay worsens the performance of OFTRL from the aspects of social regret and convergence. This study proposes the weighted OFTRL (WOFTRL), where the prediction vector of the next reward in OFTRL is weighted $n$ times. We further capture an intuition that the optimistic weight cancels out this time delay. We prove that when the optimistic weight exceeds the time delay, our WOFTRL recovers the good performances that social regret is constant in general-sum normal-form games, and the strategies last-iterate converge to the Nash equilibrium in poly-matrix zero-sum games. The theoretical results are supported and strengthened by our experiments.

Kaito Ariu, Po-An Wang, Alexandre Proutiere et al.

We study the policy testing problem in discounted Markov decision processes (MDPs) under the fixed-confidence setting. The goal is to determine whether the value of a given policy exceeds a specified threshold while minimizing the number of observations. We begin by deriving an instance-specific lower bound that any algorithm must satisfy. This lower bound is characterized as the solution to an optimization problem with non-convex constraints. We propose a policy testing algorithm inspired by this optimization problem--a common approach in pure exploration problems such as best-arm identification, where asymptotically optimal algorithms often stem from such optimization-based characterizations. As for other pure exploration tasks in MDPs, however, the non-convex constraints in the lower-bound problem present significant challenges, raising doubts about whether statistically optimal and computationally tractable algorithms can be designed. To address this, we reformulate the lower-bound problem by interchanging the roles of the objective and the constraints, yielding an alternative problem with a non-convex objective but convex constraints. Strikingly, this reformulated problem admits an interpretation as a policy optimization task in a newly constructed reversed MDP. Leveraging recent advances in policy gradient methods, we efficiently solve this problem and use it to design a policy testing algorithm that is statistically optimal--matching the instance-specific lower bound on sample complexity--while remaining computationally tractable. We validate our approach with numerical experiments.

Hiroki Ishibashi, Kenshi Abe, Atsushi Iwasaki

This paper introduces state abstraction for two-player zero-sum Markov games (TZMGs), where the payoffs for the two players are determined by the state representing the environment and their respective actions, with state transitions following Markov decision processes. For example, in games like soccer, the value of actions changes according to the state of play, and thus such games should be described as Markov games. In TZMGs, as the number of states increases, computing equilibria becomes more difficult. Therefore, we consider state abstraction, which reduces the number of states by treating multiple different states as a single state. There is a substantial body of research on finding optimal policies for Markov decision processes using state abstraction. However, in the multi-player setting, the game with state abstraction may yield different equilibrium solutions from those of the ground game. To evaluate the equilibrium solutions of the game with state abstraction, we derived bounds on the duality gap, which represents the distance from the equilibrium solutions of the ground game. Finally, we demonstrate our state abstraction with Markov Soccer, compute equilibrium policies, and examine the results.

Kenshi Abe, Kaito Ariu, Mitsuki Sakamoto et al.

This paper proposes a payoff perturbation technique for the Mirror Descent (MD) algorithm in games where the gradient of the payoff functions is monotone in the strategy profile space, potentially containing additive noise. The optimistic family of learning algorithms, exemplified by optimistic MD, successfully achieves {\it last-iterate} convergence in scenarios devoid of noise, leading the dynamics to a Nash equilibrium. A recent re-emerging trend underscores the promise of the perturbation approach, where payoff functions are perturbed based on the distance from an anchoring, or {\it slingshot}, strategy. In response, we propose {\it Adaptively Perturbed MD} (APMD), which adjusts the magnitude of the perturbation by repeatedly updating the slingshot strategy at a predefined interval. This innovation empowers us to find a Nash equilibrium of the underlying game with guaranteed rates. Empirical demonstrations affirm that our algorithm exhibits significantly accelerated convergence.

Hiroaki Shiino, Kaito Ariu, Kenshi Abe et al.

Bandit algorithms for online learning to rank (OLTR) problems often aim to maximize long-term revenue by utilizing user feedback. From a practical point of view, however, such algorithms have a high risk of hurting user experience due to their aggressive exploration. Thus, there has been a rising demand for safe exploration in recent years. One approach to safe exploration is to gradually enhance the quality of an original ranking that is already guaranteed acceptable quality. In this paper, we propose a safe OLTR algorithm that efficiently exchanges one of the items in the current ranking with an item outside the ranking (i.e., an unranked item) to perform exploration. We select an unranked item optimistically to explore based on Kullback-Leibler upper confidence bounds (KL-UCB) and safely re-rank the items including the selected one. Through experiments, we demonstrate that the proposed algorithm improves long-term regret from baselines without any safety violation.

Kenshi Abe, Junpei Komiyama, Atsushi Iwasaki

This paper considers the capacity expansion problem in two-sided matchings, where the policymaker is allowed to allocate some extra seats as well as the standard seats. In medical residency match, each hospital accepts a limited number of doctors. Such capacity constraints are typically given in advance. However, such exogenous constraints can compromise the welfare of the doctors; some popular hospitals inevitably dismiss some of their favorite doctors. Meanwhile, it is often the case that the hospitals are also benefited to accept a few extra doctors. To tackle the problem, we propose an anytime method that the upper confidence tree searches the space of capacity expansions, each of which has a resident-optimal stable assignment that the deferred acceptance method finds. Constructing a good search tree representation significantly boosts the performance of the proposed method. Our simulation shows that the proposed method identifies an almost optimal capacity expansion with a significantly smaller computational budget than exact methods based on mixed-integer programming.

Masahiro Kato, Kenshi Abe, Kaito Ariu et al.

Off-policy evaluation (OPE) is the problem of estimating the value of a target policy from samples obtained via different policies. Recently, applying OPE methods for bandit problems has garnered attention. For the theoretical guarantees of an estimator of the policy value, the OPE methods require various conditions on the target policy and policy used for generating the samples. However, existing studies did not carefully discuss the practical situation where such conditions hold, and the gap between them remains. This paper aims to show new results for bridging the gap. Based on the properties of the evaluation policy, we categorize OPE situations. Then, among practical applications, we mainly discuss the best policy selection. For the situation, we propose a meta-algorithm based on existing OPE estimators. We investigate the proposed concepts using synthetic and open real-world datasets in experiments.

Kaito Ariu, Kenshi Abe, Alexandre Proutière

In this paper, we revisit the regret minimization problem in sparse stochastic contextual linear bandits, where feature vectors may be of large dimension $d$, but where the reward function depends on a few, say $s_0\ll d$, of these features only. We present Thresholded Lasso bandit, an algorithm that (i) estimates the vector defining the reward function as well as its sparse support, i.e., significant feature elements, using the Lasso framework with thresholding, and (ii) selects an arm greedily according to this estimate projected on its support. The algorithm does not require prior knowledge of the sparsity index $s_0$ and can be parameter-free under some symmetric assumptions. For this simple algorithm, we establish non-asymptotic regret upper bounds scaling as $\mathcal{O}( \log d + \sqrt{T} )$ in general, and as $\mathcal{O}( \log d + \log T)$ under the so-called margin condition (a probabilistic condition on the separation of the arm rewards). The regret of previous algorithms scales as $\mathcal{O}( \log d + \sqrt{T \log (d T)})$ and $\mathcal{O}( \log T \log d)$ in the two settings, respectively. Through numerical experiments, we confirm that our algorithm outperforms existing methods.

Masahiro Kato, Kei Nakagawa, Kenshi Abe et al.

This study investigates the mean-variance (MV) trade-off in reinforcement learning (RL), an instance of the sequential decision-making under uncertainty. Our objective is to obtain MV-efficient policies whose means and variances are located on the Pareto efficient frontier with respect to the MV trade-off; under the condition, any increase in the expected reward would necessitate a corresponding increase in variance, and vice versa. To this end, we propose a method that trains our policy to maximize the expected quadratic utility, defined as a weighted sum of the first and second moments of the rewards obtained through our policy. We subsequently demonstrate that the maximizer indeed qualifies as an MV-efficient policy. Previous studies that employed constrained optimization to address the MV trade-off have encountered computational challenges. However, our approach is more computationally efficient as it eliminates the need for gradient estimation of variance, a contributing factor to the double sampling issue observed in existing methodologies. Through experimentation, we validate the efficacy of our approach.

Kenshi Abe, Yusuke Kaneko

Off-policy evaluation (OPE) is the problem of evaluating new policies using historical data obtained from a different policy. In the recent OPE context, most studies have focused on single-player cases, and not on multi-player cases. In this study, we propose OPE estimators constructed by the doubly robust and double reinforcement learning estimators in two-player zero-sum Markov games. The proposed estimators project exploitability that is often used as a metric for determining how close a policy profile (i.e., a tuple of policies) is to a Nash equilibrium in two-player zero-sum games. We prove the exploitability estimation error bounds for the proposed estimators. We then propose the methods to find the best candidate policy profile by selecting the policy profile that minimizes the estimated exploitability from a given policy profile class. We prove the regret bounds of the policy profiles selected by our methods. Finally, we demonstrate the effectiveness and performance of the proposed estimators through experiments.

Masahiro Nomura, Kenshi Abe

The aim of black-box optimization is to optimize an objective function within the constraints of a given evaluation budget. In this problem, it is generally assumed that the computational cost for evaluating a point is large; thus, it is important to search efficiently with as low budget as possible. Bayesian optimization is an efficient method for black-box optimization and provides exploration-exploitation trade-off by constructing a surrogate model that considers uncertainty of the objective function. However, because Bayesian optimization should construct the surrogate model for the entire search space, it does not exhibit good performance when points are not sampled sufficiently. In this study, we develop a heuristic method refining the search space for Bayesian optimization when the available evaluation budget is low. The proposed method refines a promising region by dividing the original region so that Bayesian optimization can be executed with the promising region as the initial search space. We confirm that Bayesian optimization with the proposed method outperforms Bayesian optimization alone and shows equal or better performance to two search-space division algorithms through experiments on the benchmark functions and the hyperparameter optimization of machine learning algorithms.