Recap of Lecture 15 and Homework Check
LAST LECTURE · 5 MIN
Today: the hiker at 120 m must walk DOWN first to reach 340 m. Can we allow that - but not too often?
Today's Plan (50 minutes)
ROADMAP
0-5
Recap of Lecture 15 and homework check
S02
5-8
Plan and learning outcomes
S03-S04
8-17
Module 1: Simulated annealing - idea, algorithm, acceptance probability
S05-S07
17-26
Module 2: Annealing on the track, 8-queens and a sales tour
S08-S12
26-35
Module 3: Parallel search - methods, speed-up, load balancing
S13-S16
35-43
Class activity: accept or reject? speed-up and efficiency
S17
43-48
Common mistakes and recap quiz
S18-S19
48-50
Homework, mid-term and next lecture
S20
Learning Outcomes (SLOs)
BY THE END OF THIS LECTURE YOU WILL BE ABLE TO
SLO 16.1
Explain simulated annealing and calculate the probability of accepting a worse move, e^(-drop / T)
→ CLO3 · C3 Apply
SLO 16.2
Apply simulated annealing to a landscape, 8-queens and a touring problem, and compare it with hill climbing
→ CLO3 · C3 Apply
SLO 16.3
Explain parallel search: independent runs, splitting the tree, speed-up, efficiency and load balancing
→ CLO3 · C2 Understand
- Course Learning Outcome(s) served
- CLO3: Apply various techniques and algorithms of AI used in general problem solving, optimization problems, constraint satisfaction problems, and game programming.
Slide 5
MODULE 1
Simulated Annealing
Shake hard at first, then gently
The Idea: Allow Some Downhill Steps
MODULE 1 · IDEA
Hill climbing
Random walk
Simulated annealing
Never steps down: stuck on the first peak.
Steps anywhere: finds the top by luck but never stays there.
Random steps; downhill allowed with a probability that falls as T cools.
Blacksmith: heat the iron red-hot, then cool it slowly - the metal settles into a strong state. Cool too fast and it cracks.
Temperature T: high T = adventurous (accepts bad moves); low T = careful (almost hill climbing).
The Algorithm and the Acceptance Probability
MODULE 1 · ALGORITHM
SIMULATED ANNEALING (LOCALSEARCH.PY)
REAL OUTPUT: L16/ACCEPT.PY - P = E^(-DROP/T)
- T = T0
- for step in range(max_steps):
- nxt = random neighbour of current
- drop = value(current) - value(nxt)
- if drop <= 0 or random() < exp(-drop / T):
- current = nxt # accept
- T = T * cooling # e.g. 0.995
- drop (m) T = 200 T = 100 T = 50 T = 10 T = 1
- 10 0.951 0.905 0.819 0.368 0.000
- 50 0.779 0.607 0.368 0.007 0.000
- 100 0.607 0.368 0.135 0.000 0.000
Rule: better move: always accept. Worse move: accept if a random number r < e^(-drop / T).
Example: drop 50 m at T = 100: p = e^(-0.5) = 0.607. Random number 0.4 < 0.607 - accept; 0.8 - reject.
Slide 8
MODULE 2
Annealing in Action
The track, 8-queens and a sales tour
Annealing on the Foggy Track
MODULE 2 · TRACK
REAL OUTPUT: L16/TRACK_SA.PY (T0 = 200, 1,500 STEPS)
- runs (out of 100) that end on the 340 m top
- start hill climbing SA slow (x0.995) SA fast (x0.9)
- 1 stuck at 120 m 76 / 100 0 / 100
- 5 stuck at 120 m 71 / 100 0 / 100
- 12 stuck at 250 m 78 / 100 24 / 100
Hill climbing
Slow cooling
Fast cooling
Always stuck on the nearest hill: 120 m or 250 m.
About 3 runs in 4 reach the 340 m top.
Freezes too early - behaves like hill climbing (0 of 100 from start 1).
Annealing on 8-Queens
MODULE 2 · 8-QUEENS
REAL OUTPUT: L16/QUEENS_SA.PY (100 RANDOM STARTS EACH)
- hill climbing (steepest ascent): 18 / 100 solved
- simulated annealing (T0 = 2, x0.99 per step): 96 / 100 solved, average 763 random moves tried
Lesson: a little 'bad' movement early on helps the search escape local minima.
A Real Problem: A 10-Town Sales Tour
MODULE 2 · TOURING PROBLEM
REAL OUTPUT: L16/TOUR.PY
- Exact (checks all 362,880 orders): 724 km
- Burewala - Vehari - Multan - Khanewal - Mian Channu - Sahiwal - Faisalabad - Lahore - Pakpattan - Arifwala - Burewala
- Hill climbing (2-opt) 50 random starts: best tour found 29 times, average 726.7 km, worst 730 km
- Simulated annealing 50 random starts: best tour found 44 times, average 724.9 km, worst 730 km
Touring problem
2-opt neighbour
Used for
Visit every town once and return; shortest total distance.
Reverse one section of the tour - removes crossing roads.
Delivery routes, chip layout, timetables - too big for exact search.
Properties and Tuning of Simulated Annealing
MODULE 2 · PROPERTIES
Tip: always remember the best state seen - SA can walk away from it while T is still high.
Slide 13
MODULE 3
Parallel Search
Many processors, one search
Three Ways to Search in Parallel
MODULE 3 · METHODS
Independent runs
Split the tree
Parallel expansion
Each core runs its own random restarts or SA. Keep the best. No communication.
Each core searches a different subtree (e.g. first queen in different rows).
Expand many frontier nodes at once; share the frontier; remove duplicates.
Examples: Deep Blue (1997): 30 processors + 480 chess chips. AlphaGo (2016): about 1,200 CPUs and 176 GPUs.
Link: local beam search (L15) = k runs that share their best states.
Splitting the 11-Queens Tree
MODULE 3 · DEMO
REAL OUTPUT: L16/PARALLEL.PY (TIMES VARY FROM RUN TO RUN AND COMPUTER TO COMPUTER)
- 11-queens: 2680 solutions (serial) = 2680 (parallel)
- solutions in each subtree (first queen in row 0 ... 10): [96, 219, 209, 295, 346, 350, 346, 295, 209, 219, 96]
- 1 process: 1.30 s | 2 processes: 0.78 s | speed-up 1.7x on this 2-core computer
PYTHON: GIVE SUBTREES TO WORKER PROCESSES
- from multiprocessing import Pool
- with Pool(2) as pool:
- parts = pool.map(subtree,
- [(n, r) for r in range(n)])
- total = sum(parts)
Never exactly p times: starting processes, sending work and results, and uneven parts all cost time.
Problems in Parallel Search
MODULE 3 · CHALLENGES
Subtree sizes: 96, 219, 209, 295, 346, 350, 346, 295, 209, 219, 96 - not equal, so the work must be shared out carefully.
Activity: Accept or Reject? Speed-Up and Efficiency
CLASS ACTIVITY · PAIRS · 8 MIN
Pairs, 6 min + 2 min. Use p = e^(-drop / T) (calculator allowed). Track heights: position 3 = 120 m, position 4 = 110 m.
A move drops 30 m. Find p at T = 100 and at T = 10.
Why did fast cooling (x 0.9) give 0 of 100 runs on the 340 m top from start 1?
SA is at position 3 (120 m) and tries position 4 (110 m), T = 50, random number 0.70. Accept or reject?
A search takes 12 s on 1 core and 4 s on 4 cores. Speed-up and efficiency?
Same move at T = 5, random number 0.70. Accept or reject?
Give the 11 subtrees (S16) to 4 workers in turn (1st to W1, 2nd to W2, ...). Work per worker? Is it balanced?
Common Mistakes
WATCH OUT
WRONG IDEA
CORRECT IDEA
✗ SA accepts every worse move.
✓ Only with probability e^(-drop/T) - rarer for big drops and low T.
✗ Higher temperature = more careful.
✓ High T = adventurous; low T = careful, almost hill climbing.
✗ Cooling fast is better - it is quicker.
✓ Fast cooling freezes early: 0 of 100 reached the top from start 1.
✗ p processors always give p times speed-up.
✓ Overheads and uneven work: under 2x on our 2 cores.
✗ SA returns its final state.
✓ Keep the best state seen - SA may walk away from it.
Recap Quiz
CHECK YOUR UNDERSTANDING · 3 MIN
- Q1. SA accepts a worse move with probability:
- (a) drop / T (b) e^(-drop / T) (c) T / drop (d) Always 0.5
- Q2. Drop 50 m at T = 100: p is about:
- (a) 0.37 (b) 0.50 (c) 0.61 (d) 0.95
- Q3. As the temperature T falls, SA behaves more like:
- (a) A random walk (b) Hill climbing (c) BFS (d) A*
- Q4. 12 s on 1 core and 4 s on 4 cores. Efficiency is:
- (a) 25% (b) 33% (c) 75% (d) 300%
Homework and Next Lecture
BEFORE THE NEXT CLASS
- Task 1: Calculate p = e^(-drop / T) for a 20 m drop at T = 100, 20 and 2.
- Task 2: SA on the track starts at position 3 (120 m). Positions 4 and 5 are 110 m and 80 m. Steps: try 4 (T = 100, r = 0.5); try 5 (T = 90, r = 0.9); try 5 (T = 81, r = 0.6). Where is SA after each step?
- Task 3: A search takes 40 s on 1 core and 8 s on 8 cores. Find speed-up and efficiency; give two reasons it is not 8x.
- Mid-term: Lectures 1-16. Practise the traces: BFS, DFS, UCS, greedy, A*, hill climbing, SA acceptance.
NEXT LECTURE
- After the mid-term - Lecture 17:
- Game Playing: Minimax