We develop SEGClobber (Simplest Equal Game Clobber), a solver for the one-dimensional form of Clobber known as linear Clobber. We briefly review previous work on this game and related concepts about combinatorial game theory. We describe our basic computational approach, a series of enhancements, and the experimental evaluation. We experimentally verify two conjectures by Albert et al. about linear Clobber up to values of n=50 for ( )^n and n=28 for ( )^n . This greatly exceeds previous computational results for such positions.
We propose a method to learn strong strategies in Othello, a two-player, zero-sum, finite, deterministic game with perfect information, without using existing game records or domain-specific knowledge. Silver et al. [15] proposed a self-play reinforcement learning algorithm that achieved superhuman performance in chess, Go, and shogi without human knowledge, while our method is more computationally efficient. We use N-tuple systems [3] for feature extraction and model the value function as a sigmoid applied to a linear combination of these features, yielding a predicted win rate. We optimize the value-function parameters through self-play reinforcement learning and refine the n-tuples with a genetic algorithm on self-play game records. In experiments, our method combined with Monte Carlo Tree Search (MCTS) using 800 simulations achieves a positive win rate against Edax [5], a strong open-source Othello program, at Level 5. With the αβ pruning algorithm at the same search depth, it also defeats Edax at Level 10.
Imitating human gameplay is one of the important factors in entertaining or teaching human players. In this paper, we target the game of Go at amateur levels and investigate KataGo’s HumanSL model, which was trained on a large number of human game records using supervised learning (SL). Specifically, we (1) evaluate the human-likeness of HumanSL policies (i.e., probability distribution of moves) at various skill levels from several aspects, (2) propose a method called Policy Adjustment based on Previous Move (PAPM) to enhance the human-likeness of HumanSL policies, and (3) demonstrate that selecting moves using a softmax function with appropriate parameter settings can better fit the specified skill levels. Additionally, we conduct a preliminary experimental study to compare the probability distributions of HumanSL and PAPM with softmax (PAPM+SX). The results indicate that the participants found PAPM+SX to be more human-like than HumanSL.
The goal of this study is to introduce a group of agents with diverse strategies in Go. Each agent should be strong enough (e.g., to guide human players) and still each agent have their own strategy for securing territory to win a game. For diversity, we present two variations, static and intrinsic, and evaluate them. The former enables a user to specify an area of interest to secure, e.g., center-intensive or edge-intensive strategies. The latter encourages agents to have their own strategy without user intervention. Both variations formalize their diversity objectives into pseudo rewards to balance the playing strength and diversity, suitable for the reinforcement learning of Gumbel AlphaZero. Experiments on nine-by-nine Go show both objectives yield a group of diverse agents with almost no or small drawback in playing strength depending on parameters.
Tagiron (also known as Break the Code) is a popular logic-based deduction game played with a deck of 20 cards. The rules of Tagiron vary depending on the number of players. In a four-player game, each player is dealt four cards, and the remaining four cards, called the target cards, are the goal for players to deduce. When a player makes a “challenge” to deduce the target cards, they secretly look at the hidden cards directly. If a player makes an incorrect challenge, they immediately lose and are eliminated from the game as they gain complete knowledge of the target cards. For this reason, players are limited to a single challenge throughout a game. In this paper, we propose a secure computation protocol using card-based cryptography. This protocol can determine the success or failure of a challenge without revealing the hidden target cards to any player. This innovation opens up new variations of Tagiron that allow for multiple challenges (e.g., up to two challenges), which could potentially enhance gameplay and strategic depth.
