CS 188's Personal Meeting Room - Shared screen with speaker view
Ok for me
could you go back to utility quickly?
just the last part of last slide
Slides for today: https://drive.google.com/file/d/1fxTp0EeyDJizuDYc7jjHDRs0eXQaGqAC/view?usp=sharing
this is the pdf for the slides
The states “utility” is the score associated with that state right?
we technically didnt need to explore the 4 and 6 nodes correct?
alpha beta pruning yee
does this algorithm depend on the fact that the opponent plays optimally
can you go over the successor functions?
But if you try to exploit imperfect play, doesn't that open you up to being exploited as well?
yea but you're assuming they aren't good enough to play optimally
I suppose there would be an extra component of the algorithm to weigh risk/reward against suboptimal opponents
I thought the ghosts were master minds
What's the goal test? Eat all the dots? Eat a ghost?
if the ghost is moving randomly, would the action of pacman be deterministic?
I think yeah based on movement of ghosts
In the case when ghosts are not acting optimally, are we introducing probability in the evaluation?
You could compute expected values of moves to make a choice in that case
I think the policy would be deterministic, but the actual action would random, and depend on the movement of the ghost
but since the movement of ghosts are random, how would the pacman act based on the actions of ghosts? By introducing probability?
Yeah I think then the policy would be to take the action with the best expected value
is it possible for movement of ghosts to complete random? it will be like pseudo random
That's the question of coming up with a function that generates random numbers basically, which is really hard
because how do you define random
but pseudo random yes, depending on what you mean by "random"
So how would you do that child ordering?
can you clarify how the ordering improves the algorithm?
why could the intermediate node be wrong when we use <=?
Can u give an example where ordering improves algo?
So if this is only for zero-sum games, where does it fail for non-zero-sum?
was the v <= 10 the wrong node
and 0 would've been right
what if the value of 5 was 1?
Ahh okay. So they would have to maximize each others utilities instead of minimizing their own
I think ye Estea, but what he was getting at was if there was a smaller value on the right (eg: -1) then we still get 10 as the minimax root value
scratch my question
So intermediate nodes are wrong; wouldn't that make the root node also wrong?
but then the true value should've been -1 rathre than 0
Can you please explain again why the time complexity for minimax with pruning is O(b^(m/2))
oh ok nevermind then thanks
no since we pick the 10 either way
Intermediate nodes would be wrong if they have a value already worse than the root node; the root node stays optimal
^ Danny bump
So like no matter if we picked the 0 or the -1 as the right value, the top node maximizes so it would always choose the 10 @Estea
I see.. hence the pruning?
Ty @ Calvin & Danny!
lol f in the chat
are -2 and 9 heuristics? How do we know value if we don’t reach the terminal
Yep heuristic estimates
can't we limit actions using cost
in this case what if on the board both sides don't have queens anymore? Then the heuristic will stay at 0 right?
wait how did he fix the thrashing
So what separates evaluation functions from heuristics ?
how do we know the ghosts are DFS not BFS?
what do you mean “the actual terminal utility comes into play”?
Think that means they find the actual value rather than heuristic
Fixed the thrashing with a better heuristic
better heuristic not evaluation function?
^ the evaluation of the utilities are based on proximity to the dots now, rather than seeing just the dots exist when depth = 2
from my understanding they're the same?
its pretty much the same, you use different heuristics to determine an evaluation function I think
not too sure
i see.. so they r similar
Are lecture recordings posted somewhere yet?