並列タイトル等協同的コロニーの軌道データによる先導・従エージェントとそれらの相互作⽤領域の同定に関する研究
一般注記担当:理学部図書室
Systems composed of locally interacting particles or agents such as birds, fish, and cellsshow spontaneous, spatiotemporal, collective behaviors as a whole. Exploration of theunderlying mechanisms and principles of such coordination of agents has been made inexperimental, and theoretical studies. In many systems, it has been proved that thepresence of influential individuals, known as `leaders', are responsible for the collectivemotion of agents. These leaders control the movement of the whole community.In some cases, e.g., fish shoal, MDCK epithelial cells, the relative position of theagents helps to identify the leader. But in many cases where agents do not move in thesame direction, e.g., Dictyostelium Discoideum, the relative position does not help toidentify leaders. Hence identifying leader agents in a collectively moving community is aperplexing work.Since the follower's movement is regulated by the leader agents, hence there is a correlationbetween the movement of leader and follower agents at a certain time lag. Consequently,cross-correlation has been used in identifying leader and follower agents. Butits performance is questionable in non-linear systems. Transfer entropy, an information-theoreticmeasure, is capable of capturing non-linear interaction between agents, henceit has been used to identify leader agents.The effectiveness of TE in identifying leader agents has been tested in a Dictyosteliumdiscoideum colony. It has been found that the result obtained using TE is almost identicalto the expert's result.The Vicsek model (VM) often studied as a metaphor for collectively moving animalsis employed. A modified version of the VM has helped us to investigate the classificationperformance of CC and TE. It has been found that TE outperforms CC. Different modelparameters have been varied to check their effect on classification scores.An information-theoretic scheme is proposed to estimate the underlying domain ofinteractions and the timescale of the interactions for many-particle systems. Based onensemble data of trajectories of the model system, it is shown that using the interactiondomain significantly improves the performance of classification of leaders and followerscompared to the approach without utilizing knowledge of the domain. Given an interactiontimescale estimated from an ensemble of trajectories, the first derivative of transferentropy averaged over the ensemble with respect to the cut-off distance is presented toserve as an indicator to infer the interaction domain. It is shown transfer entropy is superiorto infer the interaction radius compared to cross-correlation, hence resulting in ahigher performance to infer leader-follower relationship. Effects of noise size exerted fromthe environment, and the ratio of the numbers of leader and follower on the classificationperformance is also discussed.Later it was found that the `minimum derivative' scheme is dependent on how transferentropy can be estimated so that it takes into account enough statistics of interactingparticles, and positions and numbers of the minimum of the derivative of average transferentropy along with the cutoff distance λ may also be subject to the extent of externalnoise and time length of trajectories. The author has scrutinized how the predictionperformance in capturing the underlying interaction domain depends on the size of noiseand time length of the trajectory data. An alternative scheme has been proposed which isexpected to be stable against noises and time length, that relies on the degree of convexityat coarse-grained scale in the derivative of average transfer entropy along with the cutoffdistance, and time variance of underlying interaction radius of particles.
(主査) 教授 小松崎 民樹(電子科学研究所), 教授 芳賀 永, 教授 グン 剣萍
生命科学院(生命科学専攻)
コレクション(個別)国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文
受理日(W3CDTF)2021-07-05T22:24:43+09:00
連携機関・データベース国立国会図書館 : 国立国会図書館デジタルコレクション