Moment-Based Variational Inference for Markov Jump Processes - PowerPoint PPT Presentation

Moment-Based Variational Inference for Markov Jump Processes Christian Wildner and Heinz Koeppl Department of Electrical Engineering and Information Technology Technische Universität Darmstadt, Germany June 11, 2019 | Moment-Based Variational Inference for Markov Jump Processes | Christian Wildner, Heinz Koeppl | 1

Introduction Model Class: Markov jump process / continuous time Markov chain • Applications in many domains (finance, social networks, healthcare, systems biology, etc.) • Data-driven modelling requires latent state estimation June 11, 2019 | Moment-Based Variational Inference for Markov Jump Processes | Christian Wildner, Heinz Koeppl | 2

Introduction Model Class: Markov jump process / continuous time Markov chain • Applications in many domains (finance, social networks, healthcare, systems biology, etc.) • Data-driven modelling requires latent state estimation Problem: Hard/intractable for large state spaces June 11, 2019 | Moment-Based Variational Inference for Markov Jump Processes | Christian Wildner, Heinz Koeppl | 3

Introduction Model Class: Markov jump process / continuous time Markov chain • Applications in many domains (finance, social networks, healthcare, systems biology, etc.) • Data-driven modelling requires latent state estimation Problem: Hard/intractable for large state spaces Proposed solution: new variational inference approach based on • transition space partitioning • gradient-based optimization June 11, 2019 | Moment-Based Variational Inference for Markov Jump Processes | Christian Wildner, Heinz Koeppl | 4

Markov Jump Processes An MJP is fully defined by • an initial distribution • a transition function with June 11, 2019 | Moment-Based Variational Inference for Markov Jump Processes | Christian Wildner, Heinz Koeppl | 5

Markov Jump Processes An MJP is fully defined by • an initial distribution • a transition function with June 11, 2019 | Moment-Based Variational Inference for Markov Jump Processes | Christian Wildner, Heinz Koeppl | 6

<latexit sha1_base64="LYecnUJ+2k+l/9JOByuW6hbOWik=">ACeHicfVFNSwMxE3X7/Vbj16CpaiXslsLeix48VjB1oJdSjadtdFksySzQln6H7zqP/OveDL9QNwKDgQe781k3szEmRQWg+Cz4q2srq1vbG752zu7e/sHh0dq3PDocO1KYXMwtSpNBgRJ6mQGmYgkP8cvNVH94BWOFTu9xnEGk2FMqEsEZOqrbO8dB42JwUA3qwSzoXxAuQLVF5tEeHFb6/aHmuYIUuWTWPoZBhlHBDAouYeL3cwsZ4y/sCR4dTJkCGxUzuxNac8yQJtq4lyKdsb8rCqasHavYZSqGI7usTckfrVZqhcl1VIg0yxFSPu+U5JKiptPh6VAY4CjHDjBuhDNL+YgZxtGtyPdrlOcWtfr/05JaTF2g1tKWhx6BfAW3EZvM3JY0y5mEYVTMc0paEcdq4rt7hMvb/wu6jXp4W/cNaut5uIym+SEnJzEpIr0iK3pE06hJNn8kbeyUfly6PemXcxT/Uqi5pjUgqv8Q26icHX</latexit> <latexit sha1_base64="yliyfy+gVJBkdxTS+u+7TovaQTk=">ACenicfVFNSwMxE3X7/Vbj15CS0ERym4t6LHgxaOCtYV2Kdl01gaTzZLMCmXpn/Cqf8z/4sH0A3EtOB4vDeTeTMTZ1JYDILPire2vrG5tb3j7+7tHxweHZ8WZ0bDh2upTa9mFmQIoUOCpTQywFUvoxi+3M737CsYKnT7iJINIsedUJIzdFSvd47D5uX4YnhUCxrBPOgqCJeg1iaLuB8eVwaDkea5ghS5ZNb2wyDqGAGBZcw9Qe5hYzxF/YMfQdTpsBGxdzwlNYdM6KJNu6lSOfs74qCKWsnKnaZiuHY/tVm5I9WL7XC5CYqRJrlClfdEpySVHT2fh0JAxwlBMHGDfCmaV8zAzj6Jbk+3XKc4ta/f9pS1mLlBractDj0G+gtuITeZuS5rlTMIoKhY5Ja2IYzX13T3Cv9tfBU/NRnjVaD60au3W8jLb5IxUyTkJyTVpkztyTzqE0neyDv5qHx5Ve/Cu1ykepVlzSkphdf6BjgAwn4=</latexit> <latexit sha1_base64="YamRpAPqzTl/wPlx69cAdPmack=">ACeHicfVFNSwMxE3X7/Vbj16CpaiXslsLeix48VjB1oJdSjadtdFksySzQln6H7zqP/OveDL9QNwKDgQe781k3szEmRQWg+Cz4q2srq1vbG752zu7e/sHh0dq3PDocO1KYXMwtSpNBgRJ6mQGmYgkP8cvNVH94BWOFTu9xnEGk2FMqEsEZOqrbO8dBeDE4qAb1YBb0LwgXoNoi82gPDiv9/lDzXEGKXDJrH8Mgw6hgBgWXMPH7uYWM8Rf2BI8OpkyBjYqZ3QmtOWZIE23cS5HO2N8VBVPWjlXsMhXDkV3WpuSPViu1wuQ6KkSa5Qgpn3dKcklR0+nwdCgMcJRjBxg3wpmlfMQM4+hW5Ps1ynOLWv3/aUktpi5Qa2nLQ49AvoLbiE1mbkua5UzCMCrmOSWtiGM18d09wuXt/wXdRj28rDfumtVWc3GZTXJCTsk5CckVaZFb0iYdwskzeSPv5KPy5VHvzLuYp3qVRc0xKYX+Aa4ecHW</latexit> <latexit sha1_base64="6vCvLUNiKi5Z9kJvpabKuQatLpk=">ACenicfVFNSwMxE3X7/Wr6tFLsBQUoezWgh4LXjwqWFtol5JNZ20w2SzJbKEs/RNe9Y/5XzyYfiCugOBx3szmTczcSaFxSD4qHhr6xubW9s7/u7e/sFh9ej4yercOhwLbXpxcyCFCl0UKCEXmaAqVhCN365nevdCRgrdPqI0wixZ5TkQjO0FG93jkOw8vxbBaCxrBIuhfEK5ArU2WcT8qgwGI81zBSlyazth0GUcEMCi5h5g9yCxnjL+wZ+g6mTIGNioXhGa07ZkQTbdxLkS7YnxUFU9ZOVewyFcOx/a3NyW+tXmqFyU1UiDTLEVK+7JTkqKm8/HpSBjgKcOMG6EM0v5mBnG0S3J9+uU5xa1+v/TklrMXaDW0paHoOcgNuITRZuS5rlTMIoKpY5Ja2IYzXz3T3C39v/C56ajfCq0Xxo1dqt1W2ySk5I+ckJNekTe7IPekQTiR5JW/kvfLpnXkX3uUy1ausak5IKbzWFzXuwn0=</latexit> <latexit sha1_base64="PVtIj65BL5r68OwnfxBdsZ3WYG0=">ACe3icfVHbSgMxE3X+3rXR1+CpeCNslsVfSz4qOC1YpdSjadtaHJZklmhbL0K3zVD/NjBNML4io4EDicM5M5MxNnUlgMgo+KNze/sLi0vOKvrq1vbG5t79xbnRsOLa6lNu2YWZAihRYKlNDODAVS3iIB1dj/eEFjBU6vcNhBpFiz6lIBGfoqMf2AXbD40b/sLtVDerBJOhfEM5AtUmcdPdrnQ6Pc1zBSlyax9CoMo4IZFzCyO/kFjLGB+wZnhxMmQIbFRPHI1pzTI8m2riXIp2wPysKpqwdqthlKoZ9+1sbk9ardQKk8uoEGmWI6R82inJUVNx/PTnjDAUQ4dYNwIZ5byPjOMo9uS79cozy1q9f+nJbUYu0CtpS0P3Qf5Am4jNpm4LWmWMwm9qJjmlLQijtXId/cIf2/L7hv1MPTeuP2rNo8m1meyRfXJAQnJBmuSa3JAW4USRV/JG3iufXtU78k6mqV5lVrNLSuGdfwHEbMK5</latexit> <latexit sha1_base64="t/+ZdG62EhV5OzgpVxhcTH6eCU=">ACeHicfVFNSwMxE3Xr7p+69FLsBT1UnaroMeCF09SwdaCXUo2nbXRZLMks0JZ+h+86j/zr3gy/UBcBQcCj/dmMm9m4kwKi0HwUfGWldW16r/sbm1vbO7t5+1+rcOhwLbXpxcyCFCl0UKCEXmaAqVjCfx8NdXvX8BYodM7HGcQKfaYikRwho7q9k5wcHM62K0FjWAW9C8IF6DWIvNoD/Yq/f5Q81xBilwyax/CIMOoYAYFlzDx+7mFjPFn9gPDqZMgY2Kmd0JrTtmSBNt3EuRztifFQVT1o5V7DIVw5H9rU3Jb61eaoXJZVSINMsRUj7vlOSoqbT4elQGOAoxw4wboQzS/mIGcbRrcj365TnFrX6/9OSWkxdoNbSlocegXwBtxGbzNyWNMuZhGFUzHNKWhHauK7e4S/t/8XdJuN8KzRvD2vtc4Xl6mSQ3JETkhILkiLXJM26RBOnsgreSPvlU+Pesfe6TzVqyxqDkgpvOYX9EnB8w=</latexit> Markov Jump Processes An MJP is fully defined by • an initial distribution • a transition function with X ( t 1 ) X ( t 1 + h ) X ( t 1 + 2 h ) X ( t 2 ) X ( t 2 + h ) X ( t N ) · · · · · · discretized representation of hidden MJP June 11, 2019 | Moment-Based Variational Inference for Markov Jump Processes | Christian Wildner, Heinz Koeppl | 7