Skip to content

Latest commit

 

History

History
74 lines (59 loc) · 4.61 KB

heuristic_analysis.md

File metadata and controls

74 lines (59 loc) · 4.61 KB

Heuristic Analysis

Optimal Solutions

  air_cargo_p1     air_cargo_p2 air_cargo_p3
1 Load(C1, P1, SFO) Load(C1, P1, SFO) Load(C1, P1, SFO)
2 Fly(P1, SFO, JFK) Fly(P1, SFO, JFK) Fly(P1, SFO, ATL)
3 Unload(C1, P1, JFK) Unload(C1, P1, JFK) Load(C3, P1, ATL)
4 Load(C2, P2, JFK) Load(C2, P2, JFK) Fly(P1, ATL, JFK)
5 Fly(P2, JFK, SFO) Fly(P2, JFK, SFO) Unload(C1, P1, JFK)
6 Unload(C2, P2, SFO) Unload(C2, P2, SFO) Unload(C3, P1, JFK)
7 Load(C3, P3, ATL) Load(C2, P2, JFK)
8 Fly(P3, ATL, SFO) Fly(P2, JFK, ORD)
9 Unload(C3, P3, SFO) Load(C4, P2, ORD)
10 Fly(P2, ORD, SFO)
11 Unload(C2, P2, SFO)
12 Unload(C4, P2, SFO)

Uninformed Search

breadth_first_search air_cargo_p1 air_cargo_p2 air_cargo_p3
optimality ✅ ✅ ✅
time elapsed (s) 0.24 84.01 481.47
# expansions 43 3343 14663
# goal tests 56 4609 18098
plan length 6 9 12
depth_first_graph_search air_cargo_p1 air_cargo_p2 air_cargo_p3
optimality ❌ ❌ ❌
time elapsed 0.12 16.32 10.86
# expansions 21 624 408
# goal tests 22 625 409
plan length 20 619 392
depth_limited_search air_cargo_p1 air_cargo_p2 air_cargo_p3
optimality ❌ ❌ ➖
time elapsed 0.60 5276.41 (stoped after 3h)
# expansions 101 222719 ➖
# goal tests 271 2053741 ➖
plan length 50 50 ➖

Breadth-first search was the only uninformed algorithm that found optimal solutions for all problems, and ran in a reasonable amount of time. Depth-first search took less time and expansions to find a solution, but the plan length was considerably larger, due to the fact that DFS kept adding actions to the solution until the goal was reached (path length is almost as large as the number of expansions). The worst performing algorithm was Depth limited search, searching deep into the state space, but not deep enough as to brute-force the solution (like DFS). The solutions that DLS was able to find both have the length of 50 (the depth limit of the algorithm).

A* Search

h_ignore_preconditions air_cargo_p1 air_cargo_p2 air_cargo_p3
optimality ✅ ✅ ✅
time elapsed 0.23 32.56 124.65
# expansions 41 1421 4589
# goal tests 43 1423 4591
plan length 6 9 12
h_pg_levelsum air_cargo_p1 air_cargo_p2 air_cargo_p3
optimality ✅ ✅ ❌
time elapsed 2.10 186.18 744.12
# expansions 11 79 242
# goal tests 13 81 244
plan length 6 9 13

The ignore preconditions heuristic (being a relaxed version of the original problem, and thus - admissible [1]) was guaranteed to find the optimal solution for all problems. It was able to find a solution faster than using a planning graph and with levelsum heuristic, which did not find an optimal solution for all problems (the heuristic is not admissible [1]). However, using levelsum resulted in considerably fewer expansions, but the time spent computing the heuristic for each expansion cancels that.

Conclusions

The heuristic that yielded the best results was ignore preconditions. Compared to non-heuristic search, it expanded less nodes than BFS, but still found optimal solutions for each problem (compared to DFS). While planning graph and with levelsum heuristic solution for the third problem was not optimal, it was very close (1 extra action on solution) while expanding fewer nodes than any other search algorithm.

  1. Stuart J. Russell and Peter Norvig (2009) "Artificial Intelligence: A Modern Approach" Prentice Hall