CS 730/830: Intro AI 1 handout: slides Control Wheeler Ruml (UNH) - PowerPoint PPT Presentation

CS 730/830: Intro AI 1 handout: slides Control Wheeler Ruml (UNH) Lecture 6, CS 730 – 1 / 12

EOLQs Control Wheeler Ruml (UNH) Lecture 6, CS 730 – 2 / 12

Control ■ Problems ■ MPC ■ Break ■ P Control ■ PD Control ■ PID Control ■ Bisection Search ■ See Also Control ■ EOLQs Wheeler Ruml (UNH) Lecture 6, CS 730 – 3 / 12

Planning Problems Observability: complete, partial, hidden Control State: discrete, continuous ■ Problems ■ MPC Actions: deterministic, stochastic, discrete, continuous ■ Break ■ P Control Nature: static, deterministic, stochastic ■ PD Control Interaction: one decision, sequential ■ PID Control ■ Bisection Search Time: static/off-line, on-line, discrete, continuous ■ See Also Percepts: discrete, continuous, uncertain ■ EOLQs Others: solo, cooperative, competitive Wheeler Ruml (UNH) Lecture 6, CS 730 – 4 / 12

Model Predictive Control used with ‘receeding horizon’ ( ≈ real-time search) Control ■ Problems ■ MPC ■ Break simulate a bunch of controls (near nominal), pick best! ■ P Control ■ PD Control ■ PID Control ■ Bisection Search or steer to a bunch of states (near nominal), pick best! ■ See Also ■ EOLQs flexible, dangerous Wheeler Ruml (UNH) Lecture 6, CS 730 – 5 / 12

Break asst3 ■ Control projects ■ Problems ■ ■ MPC wildcard class ■ ■ Break ■ P Control ■ PD Control ■ PID Control ■ Bisection Search ■ See Also ■ EOLQs Wheeler Ruml (UNH) Lecture 6, CS 730 – 6 / 12

P Control Control ■ Problems u = K P ( x r − ˆ x ) ■ MPC ■ Break ■ P Control responsiveness vs smoothness ■ PD Control = spring model ■ PID Control ■ Bisection Search unstable with inertia! ■ See Also ■ EOLQs Wheeler Ruml (UNH) Lecture 6, CS 730 – 7 / 12

PD Control Control d ( x r − ˆ x ) ■ Problems u = K P ( x r − ˆ x ) + K D ■ MPC dt ■ Break ■ P Control dampen correction if error is changing a lot ■ PD Control ■ PID Control = dampened spring model! ■ Bisection Search does nothing if persistent error balances P component ■ See Also ■ EOLQs Wheeler Ruml (UNH) Lecture 6, CS 730 – 8 / 12

PID Control Control d ( x r − ˆ x ) � ■ Problems u = K P ( x r − ˆ x ) + K I ( x r − ˆ x ) dt + K D ■ MPC dt ■ Break ■ P Control removes any persistent error ■ PD Control ■ PID Control however, ‘wind-up’ ■ Bisection Search ■ See Also ■ EOLQs widely used. not optimal or necessarily stable. tune by hand, or Thrun says coordinate-wise bisection search Wheeler Ruml (UNH) Lecture 6, CS 730 – 9 / 12

Bisection Search given f and initial guesses l and r Control ■ Problems 1. bracket a local minimum ■ MPC ■ Break (a) try guess m in middle ■ P Control ■ PD Control (b) if m smallest, done! (local min between l and r ) ■ PID Control (c) if l smallest, r ← m , m ← l and move l left ■ Bisection Search ■ See Also move l by at least original r − l (double interval) ■ EOLQs (d) if r smallest, m ← r and move r right 2. refine estimate (a) try lm between l and m . (b) if smaller than m , r ← m and m ← lm (c) otherwise, try mr between m and r . (d) if smaller than m , l ← m and m ← mr (e) otherwise m is smallest, l ← lm and r ← mr (f) until range small or values close Wheeler Ruml (UNH) Lecture 6, CS 730 – 10 / 12

See Also optimal control: eg, Linear-Quadratic-Gaussian (LQG) Control discrete control: eg, Markov decision processes ■ Problems ■ MPC state estimation aka filtering: eg, Kalman filter, particle filter ■ Break ■ P Control ■ PD Control ■ PID Control ■ Bisection Search ■ See Also ■ EOLQs Wheeler Ruml (UNH) Lecture 6, CS 730 – 11 / 12

EOLQs Please write down the most pressing question you have about Control the course material covered so far and put it in the box on your ■ Problems ■ MPC way out. ■ Break ■ P Control Thanks! ■ PD Control ■ PID Control ■ Bisection Search ■ See Also ■ EOLQs Wheeler Ruml (UNH) Lecture 6, CS 730 – 12 / 12