Automated generation of Manic Miner speedrun/walkthrough

[IRF]

On 7/24/2023 at 1:30 PM, DigitalDuck said:

About time I made an account here, I think...

So crem initially made these to help with speedrunning the game, and they were very useful for us in either confirming that our routes were optimal or finding new routes.

I took crem's inputs and used them (in a modified form) to help produce a fully optimised TAS (tool-assisted speedrun; essentially completing the game as quickly as possible using tools like slowdown, frame advance, rewind/rollback, savestates etc.) which you can find here: https://tasvideos.org/5052M

Anyway, in the progress of doing so I managed to save a single frame (i.e. score an additional point) on Endorian Forest - scoring 1841 points. This was due to some forward planning on the collapsing ground - I spend as much time as possible on it while climbing up and it results in one less frame waiting for it to disappear on the way back.

This means the highest obtainable score for the Bug-Byte version is 39531 and the highest obtainable score for the Software Projects version is 39532.

Which platform specifically are you talking about? Could you describe it please, or show an image with it circled or something? 

[IRF]

Actually, I've just added a timestamp in a comment on your YouTube video (8:52), just before when Will jumps onto the crumbly platform which I think is the one you're talking about (the lower one of the right). Is that correct?

[IRF]

I wonder how the AI failed to pick it up? Is it because it involves momentary inaction in the present to make things more efficient in the future? Mind you, that's the sort of things that humans are less likely to come up with (our in-built instinct is to do something), whereas the brute force AI approach of trying every possible iteration of moves (or in this case non-moves) should have detected it?

[IRF]

2 hours ago, IRF said:

Actually, I've just added a timestamp in a comment on your YouTube video (8:52), just before when Will jumps onto the crumbly platform which I think is the one you're talking about (the lower one of the right). Is that correct?

To answer my own question, no the difference occurs on the top right crumbly platform - Willy walks along one more step before jumping rightwards off that platform in Digital Duck's video, compared with Crem's AI-generated walk through in page 6 of this thread. However, they both land in the same place after the jump (in the optimum position to collect the top-right it), because of different left-right keypresses being pressed at the moment of landing (I think?)

Edited August 24, 2023 by IRF

[IRF]

Digital Duck took advantage of Willy's ability, when jumping up and landing on a platform above that which he jumped off, to perform a turnaround and vertical jump in a single time frame.

Crem's AI did those manoeuvres (in order to collect the top right item) in two consecutive time frames (also had to keep the right key pressed at the end of the first jump to advance Willy by one more frame to be able to collect the item using a vertical jump.

So whilst the AI was seemingly 'one step ahead' jumping up off the crumbly, that initial gain was negated by having to take two frames to turn around and jump compared to DD's one frame - at that point they're level-pegging.

But then DD's extra frame spent on the upper-right crumbly meant that they didn't have to wait as long (one time frame less) to drop through it later.  The middle of the three crumbly blocks is the determinant here - DD has things optimised as both left and middle crumbles disappear at exactly the same time, whereas in the AI recording, once the left-hand crumbly block has gone there is still one more pixel-row of the middle crumbly block remaining to delay Willy's drop.

[jetsetdanny]

Your meticulous analyses never cease to amaze me, Ian! 🙂 Thank you for this explanation. I wanted to ask the same question - why did the AI fail to detect the best sequence of movements?, but didn't, due to shortage of time. In any case, I wouldn't have analysed the issue myself like you did - wouldn't have even tried. So thanks again!

In view of your conclusion, do you think there may be other instances in the game when improvements are still possible?

[IRF]

3 hours ago, jetsetdanny said:

Your meticulous analyses never cease to amaze me, Ian! 🙂 Thank you for this explanation. I wanted to ask the same question - why did the AI fail to detect the best sequence of movements?, but didn't, due to shortage of time. In any case, I wouldn't have analysed the issue myself like you did - wouldn't have even tried. So thanks again!

In view of your conclusion, do you think there may be other instances in the game when improvements are still possible?

I've only analysed The Endorian Forest today. I guess other rooms where Willy has to progress downwards by waiting for crumblies to disappear might yield further results?

I did wonder why, in both DD and Crem's AI recordings, Willy walked out an extra step (than seemed the bare minimum) onto the crumbly platforms on the way down near the end of Endorian Forest. But then I remembered how you have to step back a bit from a fire cell before trying to jump over it. So there's no further efficiency to be found there, due to the spiky leaf fire cell near the bottom just to the right of the trunk.

(In fact, attempting a quick testing out of that manoeuvre (based on a snapshot in QAOPs that I pulled up for convenience, which happened to be from Norman Sword's Manic Jet Set Willy) is what led me to discover the slight difference between the MJSW and MM game engines, as I reported elsewhere.)

Edited August 24, 2023 by IRF

[jetsetdanny]

Thank you, Ian 🙂 

[IRF]

33 minutes ago, IRF said:

I've only analysed The Endorian Forest today. I guess other rooms where Willy has to progress downwards by waiting for crumblies to disappear might yield further results?

Although now that I think about it, I've already done an in-depth analysis of how to descend through a field of crumblies (a technique which is needed most notably in The Vat and The Warehouse in order to maximise the score). It even formed the basis of one of my mini-puzzles that you referred to earlier (funny how all these threads are coming together, isn't it?)

The Endorian Forest is, as far as I can tell, the only MM cavern where Willy has to use a crumbly platform to ascend the cavern and then use the same crumbly platform to descend back down again*. That's the scenario in which the efficiency which DigitalDuck discovered was to be found. So perhaps it is a one-off?

(*I'm not including cases where Willy descends down through an empty gap where a crumbly platform used to be; in this case the fall would be fatal if Willy dropped through it without partially crumbled platforms to stagger his fall.)

***

Anyway, it's nice to know that humans (or even ducks!?) can still outwit computers in this Age of AI... 

[IRF]

I wonder if someone updated the keypress datastream for The Endorian Forest (I could have a go myself although I'm about to fly off to Nice), whether @NormanSword might consider updating his game file which consisted of automated optimised routes through the caverns, to take account of this new discovery?