本研究では，従来，移動体通信では使用が困難と考えられていたミリ波を用いた移動体通信の実現を目指して研究した．ミリ波のような高周波では，建物や人体による遮蔽により通信が容易に遮断されることが問題となる．そのような遮蔽は，例えば建物であれば，場所によって遮蔽しやすい場所などがあり，また，人体による遮蔽であれば，人が多い場所や時間などがある．すなわち場所や時間に大きく依存することがわかる．そこで，まず，各エリアでの端末との接続状況や各端末の移動状況といった空間統計情報と地図データなどから，各エリアでの遮蔽の起こりやすさを推定および学習するアルゴリズムを開発した．データベースとして，遮蔽の起こりやすさなどの情報を持っている場合には，それを実際の接続状況を用いて，オンラインで更新する．通常，強化学習のように，ランダムな初期状態から接続による経験を蓄積していく場合は，環境に対する十分な空間統計情報を得るのに時間がかかる．本研究では，その際の収束状態に達するまでの時間を考慮した複数基地局からのビーム選択アルゴリズムを提案・評価した．また，その際，近年注目されている確率幾何に基づき遮蔽確率を推定す式を導出した．提案アルゴリズムでは，導出した遮蔽確率と学習アルゴリズムの中の強化学習アルゴリズムである多腕バンディットアルゴリズムに基づくビーム選択法を提案した．計算機シミュレーションにより提案法による接続確率と伝送レート特性を，複数基地局が存在する環境で，LTE (Long Term Evolution)のフレームフォーマットに基づき評価した．その結果，提案法は，従来の瞬時電波強度 (RSSI: Received Signal Strength Indicator)に基づくビーム選択法に比べ，特性が所要品質を下回るアウテージ確率を大幅に低減できることを示した．それらの成果を，複数の国内学会で発表した．また，国際会議・論文誌への投稿も4月中に行う予定である．
In this research, we aimed at the realization of mobile communication using millimeter waves, which was considered difficult to use in mobile communication. At high frequencies such as millimeter waves, it is a problem that communication is easily interrupted due to blocking by buildings and human bodies. For example, in the case of a building, such blocking may be a place where blocking is easy depending on the place, and in the case of blocking by a human body, there may be a place or time when there are many people. In other words, it depends on the place and time. Therefore, we developed an algorithm to estimate and learn the likelihood of blocking in each area from spatial statistical information such as connection status with terminals in each area and movement status of each terminal and map data. If you have information such as the probability of blocking as a database, update it online using the actual connection status. Usually, when accumulating experience by connection from random initial state like reinforcement learning, it takes time to obtain sufficient spatial statistical information for the environment. In this study, we proposed and evaluated beam selection algorithms from multiple base stations considering the time to reach the convergence state. At the same time, we derived an equation to estimate the occlusion probability based on the stochastic geometry that has attracted attention in recent years. The proposed algorithm proposes a beam selection method based on the derived occlusion probability and the multi-arm bandit algorithm, which is a reinforcement learning algorithm in the learning algorithm. The connection probability and transmission rate characteristics by the proposed method were evaluated by computer simulation based on the frame format of LTE (Long Term Evolution) in an environment with multiple base stations. As a result, it is shown that the proposed method can significantly reduce the outage probability whose quality is less than the required quality compared to the conventional beam selection method based on the Received Signal Strength Indicator (RSSI). The results were presented at several domestic conferences. We will also submit to international conferences and journals during April.