BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York BEGIN:DAYLIGHT DTSTART:20240310T030000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:EDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20231105T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:EST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20231220T150111Z UID:2D5C88F9-C57E-4DD5-B004-6F40CF798D4B DTSTART;TZID=America/New_York:20231213T130000 DTEND;TZID=America/New_York:20231213T140000 DESCRIPTION:We will discuss progress on two minimax problems in reinforceme nt learning. The first is a standard Markov game\, where two agents are co mpeting for rewards in an environment described by a Markov decision proce ss. Given a description of this environment\, our goal is to compute a Nas h equilibrium\, a pair of policies that neither agent can improve on in re sponse to the other. We show that a simple gradient descent/ascent algorit hm converges linearly for a regularized version of this problem\, and we c an acheive sublinear convergence to the true equilibrium through careful m anipulation of the regularization parameter.\n\nIn the second part of the talk\, we will discuss a standard form of robust reinforcement learning wh ere a (single) agent searches for a policy that has the best worst-case pe rformance over a convex set of reward functions. Although this problem is convex in one variable but non-concave in the other\, we show that there i s minimax equality. Key to this result is showing that the superlevel sets of the long-term reward for the agent are connected. We also show that th e minimax result extends to policies that are parameterized with neural ne tworks.\n\nDr. Romberg is also presenting another talk at 6pm [Dimensional ity Reduction For Sensor Arrays](https://events.vtools.ieee.org/m/384377)\ n\nSpeaker(s): Justin Romberg\n\nRoom: CST 4-201\, Bldg: Center of Science & Technology\, Syracuse University\, 111 College Pl\, Syracuse\, New York \, United States\, 13210\, Virtual: https://events.vtools.ieee.org/m/38437 2 LOCATION:Room: CST 4-201\, Bldg: Center of Science & Technology\, Syracuse University\, 111 College Pl\, Syracuse\, New York\, United States\, 13210\ , Virtual: https://events.vtools.ieee.org/m/384372 ORGANIZER:stone@ieee.org SEQUENCE:14 SUMMARY:Minimax Problems in Reinforcement Learning URL;VALUE=URI:https://events.vtools.ieee.org/m/384372 X-ALT-DESC:Description: <br /><p style="font-weight: 400\;">We will discuss progress on two minimax problems in reinforcement learning.&nbsp\; The fi rst is a standard Markov game\, where two agents are competing for rewards in an environment described by a Markov decision process.&nbsp\; Given a description of this environment\, our goal is to compute a Nash equilibriu m\, a pair of policies that neither agent can improve on in response to th e other.&nbsp\; We show that a simple gradient descent/ascent algorithm co nverges linearly for a regularized version of this problem\, and we can ac heive sublinear convergence to the true equilibrium through careful manipu lation of the regularization parameter.</p>\n<p style="font-weight: 400\;" >&nbsp\;</p>\n<p style="font-weight: 400\;">In the second part of the talk \, we will discuss a standard form of robust reinforcement learning where a (single) agent searches for a policy that has the best worst-case perfor mance over a convex set of reward functions.&nbsp\; Although this problem is convex in one variable but non-concave in the other\, we show that ther e is minimax equality.&nbsp\; Key to this result is showing that the super level sets of the long-term reward for the agent are connected.&nbsp\; We also show that the minimax result extends to policies that are parameteriz ed with neural networks.</p>\n<p style="font-weight: 400\;">&nbsp\;</p>\n< p style="font-weight: 400\;">Dr. Romberg is also presenting another talk a t 6pm <a href="https://events.vtools.ieee.org/m/384377" target="_blank" re l="noopener">Dimensionality Reduction For Sensor Arrays</a></p> END:VEVENT END:VCALENDAR