Tomoyuki Kaneko

Shanchuan Wan, Yujin Tang, Yingtao Tian et al.

Exploration is a fundamental aspect of reinforcement learning (RL), and its effectiveness is a deciding factor in the performance of RL algorithms, especially when facing sparse extrinsic rewards. Recent studies have shown the effectiveness of encouraging exploration with intrinsic rewards estimated from novelties in observations. However, there is a gap between the novelty of an observation and an exploration, as both the stochasticity in the environment and the agent's behavior may affect the observation. To evaluate exploratory behaviors accurately, we propose DEIR, a novel method in which we theoretically derive an intrinsic reward with a conditional mutual information term that principally scales with the novelty contributed by agent explorations, and then implement the reward with a discriminative forward model. Extensive experiments on both standard and advanced exploration tasks in MiniGrid show that DEIR quickly learns a better policy than the baselines. Our evaluations on ProcGen demonstrate both the generalization capability and the general applicability of our intrinsic reward. Our source code is available at https://github.com/swan-utokyo/deir.

Tomoyuki Kaneko, Shuhei Yamashita

2048 is a stochastic single-player game involving 16 cells on a 4 by 4 grid, where a player chooses a direction among up, down, left, and right to obtain a score by merging two tiles with the same number located in neighboring cells along the chosen direction. This paper presents that a variant 2048-4x3 12 cells on a 4 by 3 board, one row smaller than the original, has been strongly solved. In this variant, the expected score achieved by an optimal strategy is about $50724.26$ for the most common initial states: ones with two tiles of number 2. The numbers of reachable states and afterstates are identified to be $1,152,817,492,752$ and $739,648,886,170$, respectively. The key technique is to partition state space by the sum of tile numbers on a board, which we call the age of a state. An age is invariant between a state and its successive afterstate after any valid action and is increased two or four by stochastic response from the environment. Therefore, we can partition state space by ages and enumerate all (after)states of an age depending only on states with the recent ages. Similarly, we can identify (after)state values by going along with ages in decreasing order.

Masayo Tomita, Katsuhiko Hayashi, Tomoyuki Kaneko

Vision-Language Models (VLMs) occasionally generate outputs that contradict input images, constraining their reliability in real-world applications. While visual prompting is reported to suppress hallucinations by augmenting prompts with relevant area inside an image, the effectiveness in terms of the area remains uncertain. This study analyzes success and failure cases of Attention-driven visual prompting in object hallucination, revealing that preserving background context is crucial for mitigating object hallucination.

Yuji Kanagawa, Tomoyuki Kaneko

We consider the problem of autonomously learning reusable temporally extended actions, or options, in reinforcement learning. While options can speed up transfer learning by serving as reusable building blocks, learning reusable options for unknown task distribution remains challenging. Motivated by the recent success of mutual information (MI) based skill learning, we hypothesize that more diverse options are more reusable. To this end, we propose a method for learning termination conditions of options by maximizing MI between options and corresponding state transitions. We derive a scalable approximation of this MI maximization via gradient ascent, yielding the InfoMax Termination Critic (IMTC) algorithm. Our experiments demonstrate that IMTC significantly improves the diversity of learned options without extrinsic rewards combined with an intrinsic option learning method. Moreover, we test the reusability of learned options by transferring options into various tasks, confirming that IMTC helps quick adaptation, especially in complex domains where an agent needs to manipulate objects.

Quentin Gendre, Tomoyuki Kaneko

Catan is a strategic board game having interesting properties, including multi-player, imperfect information, stochastic, complex state space structure (hexagonal board where each vertex, edge and face has its own features, cards for each player, etc), and a large action space (including negotiation). Therefore, it is challenging to build AI agents by Reinforcement Learning (RL for short), without domain knowledge nor heuristics. In this paper, we introduce cross-dimensional neural networks to handle a mixture of information sources and a wide variety of outputs, and empirically demonstrate that the network dramatically improves RL in Catan. We also show that, for the first time, a RL agent can outperform jsettler, the best heuristic agent available.

Yuji Kanagawa, Tomoyuki Kaneko

In this paper, we propose Rogue-Gym, a simple and classic style roguelike game built for evaluating generalization in reinforcement learning (RL). Combined with the recent progress of deep neural networks, RL has successfully trained human-level agents without human knowledge in many games such as those for Atari 2600. However, it has been pointed out that agents trained with RL methods often overfit the training environment, and they work poorly in slightly different environments. To investigate this problem, some research environments with procedural content generation have been proposed. Following these studies, we propose the use of roguelikes as a benchmark for evaluating the generalization ability of RL agents. In our Rogue-Gym, agents need to explore dungeons that are structured differently each time they start a new game. Thanks to the very diverse structures of the dungeons, we believe that the generalization benchmark of Rogue-Gym is sufficiently fair. In our experiments, we evaluate a standard reinforcement learning method, PPO, with and without enhancements for generalization. The results show that some enhancements believed to be effective fail to mitigate the overfitting in Rogue-Gym, although others slightly improve the generalization ability.