-
Notifications
You must be signed in to change notification settings - Fork 0
Lecture & Tutorial Notes
vv111y edited this page Dec 28, 2015
·
1 revision
####TUT Nov20 CHESS PROJECT
include experiment protocols (same what I did in asgn2 parameter section)
Notification/warning if you are in check
4-ply minimum - MAJOR MARKS DO NOT SKIMP - turns should be around 15sec
have GUI
indicate whose turn it is
ability to restart game after losing
report which details your creative heuristic - GO IN DEPTH, MORE THE BETTER
ME
conditionals
what moves can be made testing
condition if in check
condition if can do special moves
####LEC Oct26,28
Games (Adverserial Search, also PROJECT)
heuristics definitely needed since search space so big in most games
pseudo code slide - NB* , refer to it. Project will work perfectly if you do this slide.
alpha-beta pruning (see slide examples)
you are always picking best move for yourself. don't need to expand moves you will not make
assume opponent will make best, don't expand ones he will do worse
pseudo algo slide : wrapper fn first, calls max fn, then
max fn calls min, then min calls max, ... they call each other
global ply count is tracked within this
'perfect ordering' function of rules of chess, otherwise ie. random
####LEC11 Oct28-15 CONT minimax search resource limits SOL cut off search early eg. depth limited heuristic eval fn (use EVAL instead of UTILITY)
CUTOFF
select
horizon effect
eval fn slide - [BONUS][not recommended, be creative and make something better], easy but...
linear weighted sum of features one
checkmate - harder to handle - flag allowed moves when in check or going into check
[BONUS how innovative is heuristic fn is a big chunk of marks]
if your fn is doing *consistently* worse than niave weight fn, will lose marks
but occasional dumb lossess ok
howto eval your AI? play games with it. could try playing someone elses AI
[Get basic minimax functionality first]
[look online for info on implement] [no GUI, not much marks] [AI should beat you most of the time]
GAMES WITH CHANCE
chance nodes - don't know outcomes
can use uniform distr when no info
called expectiminimax
search complexity much higher now. ply limited to about 2
often good to not even use minimax in these cases
talks of history of AI in games and current situation : Go is current game challenge, hard problem