Á¤º¸°úÇÐȸ³í¹®Áö (Journal of KIISE)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
»çÀ̹ö¹°¸® ½Ã½ºÅÛÀÇ ¾ÈÀüÇÑ °ÈÇнÀÀ» À§ÇÑ ¾ÈÀü°¡µå¿Í °¡»ó°æÇèÁÖÀÔ ±â¹ý |
¿µ¹®Á¦¸ñ(English Title) |
Safety Guards and Virtual Experience Injection Techniques for Safe Reinforcement Learning of Cyber-Physical Systems |
ÀúÀÚ(Author) |
±è¿µÀç
È«ÀåÀÇ
Youngjae Kim
Jang-Eui Hong
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 49 NO. 02 PP. 0145 ~ 0156 (2022. 02) |
Çѱ۳»¿ë (Korean Abstract) |
Çö½Ç¼¼°è¿Í °¡»ó¼¼°è¸¦ ¿¬°áÇÏ´Â CPS(Cyber-Physical System)´Â ´Ù¾çÇÑ ºÐ¾ß¿¡¼ È°¿ëµÈ´Ù. ÇÑÆí CPS¿Í ÀΰøÁö´ÉÀÇ ÇÑ ºÐ¾ßÀÎ °ÈÇнÀÀÇ µµÀÔÀº ÃÖ±Ù ¿¬±¸ÀÇ °ü½É»çÀÌ´Ù. ±×·¯³ª °ÈÇнÀ ƯÀ¯ÀÇ Å½»ö °úÁ¤¿¡¼ ¹ß»ýÇÏ´Â ¹«ÀÛÀ§¼ºÀº ¾ÈÀüÇʼöÀÎ CPS¸¦ À§ÇèÇÑ »óÅ·ΠÀüÀ̽Ãų ¼ö ÀÖ´Ù. º» ³í¹®¿¡¼´Â CPSÀÇ ¾ÈÀüÇÑ °ÈÇнÀÀ» À§ÇÑ ¾ÈÀü°¡µå¿Í °¡»ó°æÇèÁÖÀÔ ±â¹ýÀ» Á¦½ÃÇÑ´Ù. ¾ÈÀü°¡µå´Â CPS°¡ ÇнÀ µµÁß À§ÇèÇÑ »óÅ·ΠÀüÀÌÇÏ´Â °ÍÀ» ¹æÁöÇÏÁö¸¸ À§ÇèÇÑ »óÅÂÀÇ ÇнÀ °æÇèÀ» °®Áö ¾Ê°Ô ÇÑ´Ù´Â ´ÜÁ¡À» °®´Â´Ù. ÀÌ·¯ÇÑ ´ÜÁ¡Àº À§Çè »óÅ¿¡¼ÀÇ °¡»ó °æÇèÀ» ÇнÀ °úÁ¤¿¡ ÁÖÀÔÇÏ´Â °¡»ó°æÇè ÁÖÀÔÀ» ÅëÇØ ÃÖ¼ÒȽÃŲ´Ù. Á¦½ÃµÈ ¹æ¹ýÀº CPSÀÇ ¾ÈÀüÇÑ °ÈÇнÀÀ» º¸ÀåÇϸç, À§Çè »óÅ·ΠÀüÀÌµÈ °æ¿ì¿¡µµ ¾ÈÀüÇÑ »óÅ·Πº¹±ÍÇÒ ¼ö ÀÖ´Â ÀÏÂ÷ÀûÀÎ ¾ÈÀü¸ÁÀ» Á¦°øÇØÁØ´Ù. ¶ÇÇÑ ½Ã¹Ä·¹À̼ÇÀ» ÅëÇØ ¿¬±¸ °á°úÀÇ È¿¿ë¼ºÀ» ÀÔÁõÇÏ¿´´Ù. |
¿µ¹®³»¿ë (English Abstract) |
A Cyber-Physical System(CPS) that connects the real world and the cyber world is increasing in its application in diverse areas. Among the research on artificial intelligence, reinforcement learning, in particular, is achieving higher processing performance by learning the optimal policy with taking the reward. The convergence of reinforcement learning and CPS has been the focus of recent research. However, the randomness arising from the exploration by reinforcement learning can cause the problem of being able to transit safety-critical CPS to a dangerous state. This paper attempts to support the safe operation of CPS by proposing safety guards and virtual experience injection techniques for safe reinforcement learning of CPS. Although a safety guard prevents the CPS from transitioning to a dangerous state during learning, the guard has a disadvantage as it does not have a learning experience for the dangerous state. Virtual experience injection can minimize this disadvantage for a dangerous state into the learning process. The proposed safety guard and virtual experience injection techniques provide a primary safety device for transitioning to a safe state instead of a dangerous state while ensuring safe reinforcement learning of CPS. This approach has proven its effectiveness through an experimental study and simulations. |
Å°¿öµå(Keyword) |
»çÀ̹ö¹°¸® ½Ã½ºÅÛ(CPS)
°ÈÇнÀ
¾ÈÀü°¡µå
°¡»ó°æÇèÁÖÀÔ
¼ÒÇÁÆ®¿þ¾î ¾ÈÀü¼º
Cyber-Physical Systems
reinforcement learning
safety guard
virtual experience injection
software safety
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|