Mass Effect 2 Final Mission Analysis

This document assumes the reader is familiar with the entirety of Mass Effect 2, including references and terminology that may be considered spoilers. You have been warned! :)

This Rust crate defines types that encode decision paths and outcomes related to the final mission of Mass Effect 2. Although the game presents many, many choices to the player, only certain ones affect the survival of your allies into and through Mass Effect 3, if you transfer your save file. This crate is only concerned with those aspects of the game.

Outcome Data

The data generated by the generate::outcome_map() function is a map of outcomes to metadata describing the decision paths that lead to those outcomes. The metadata contains one example of a decision path and a count of the total number of decision paths that lead to the same outcome, rather than storing all of the decision paths—there are over 64 billion of them!

An Outcome encodes the following information:

• Which allies survived?
• Which surviving allies were loyal?
  • An ally becomes loyal if the player successfully completes their loyalty mission.
• Was the surviving crew of The Normandy SR2 rescued?
  • "The surviving crew" refers to the group that is still alive when Shepard arrives to rescue them.

A DecisionPath encodes the choices that affect the outcome. The examples included in each outcome's metadata answer the following questions, some of which are optional, depending on previous choices:

• Which optional allies should be recruited?
• Which loyalty missions should be completed?
• Which upgrades for The Normandy should be purchased?
• Who should be selected for the squad to defend The Normandy's cargo bay?
• Who should be selected for the various specialist roles?
• Who should be selected for the squad in the biotic shield?
• Who should be selected for the squad in the final battle?

Some player actions for which consequences for allies are realized in Mass Effect 2 and beyond are not considered in the scope of this project. See Limitations for a discussion of the rationale behind the exclusions.

To generate the data yourself, run the generate example provided in the source.

$ cargo run --release --features generate --example generate -- PATH

Alternatively, you can reuse the data included in outcome_map.rmp. It is an OutcomeMap serialized in the MessagePack format (via the rmp-serde crate).

Interesting Facts

To show the kind of questions that can be answered with the data, the following statistics were generated with the provided analyze example (and re-formatted for markdown).

$ cargo run --example analyze -- outcome_map.rmp

• There are 64,396,302,636 total decision paths.
• There are 714,852 total outcomes.
• 111 outcomes are uniquely achievable.
  • These outcomes have the following in common:
    • Miranda survives and is loyal.
    • Garrus and Jacob are disloyal.
• The most common outcome can be achieved in 68,263,592 ways:
  • Only Jacob and Miranda survive, and both are loyal.
  • The crew is rescued.
• The 10 most common outcomes cover 419,652,725 decision paths.
• The 100 most common outcomes cover 2,068,928,184 decision paths.
• The 1000 most common outcomes cover 8,572,525,662 decision paths.
• Shepard dies in only 43 outcomes from 341,570,226 decision paths.
  • In other words, Shepard survives 99.47% of all decision paths.
• The "best" possible outcome (a subjective measure, to be fair), in which everyone (except for Morinth) survives and is loyal, can be achieved in 7,968 ways.
  • If you managed to do it on your first run, give yourself a pat on the back!

Survival Rates

The following table, which shows the absolute survival rates for each ally under different conditions, is also generated by the analyze example. Allies are sorted in descending order by their total absolute survival rate.

NOTE: Survival implies recruitment. That is, if an ally is not recruited, they are not regarded as surviving.

Limitations

The following subsections discuss some of the known and perceived limitations of this crate.

Partial Data

As previously mentioned, only as many decision paths are recorded in the data as there are outcomes even though there is a one-to-many relationship between them.

As a quick thought experiment, suppose each decision path could be somehow perfectly encoded and compressed into four-and-a-half bytes (36 bits). The decision paths alone would require almost 290 GB of memory/storage—and that's the best case!

In reality, the encoding would require more than 36 bits, so it's not reasonable to expect potential consumers of this crate to sacrifice so much of their storage to host the data. By embracing this limitation, it is possible to provide the fully-generated data within the crate's source repository so users don't have to spend the 10–15 minutes it would take to generate it themselves.

Scope

The scope of this project is limited to only the parameters that affect the fate of Mass Effect 2 allies when carried over to Mass Effect 3. If an ally survives ME2, they will be encounterable in ME3. Furthermore, if they are loyal, they may become a war asset in ME3. If they were not loyal, however, they would die in ME3.

However, there are two members of the crew of The Normandy SR2 who are not specifically addressed in this implementation: Dr. Karen Chakwas and YN Kelly Chambers. Outcomes only encode whether the crew was rescued, but that is only part of the story. Some of the crew may die before they are rescued based on the number of missions completed after the installation of the Reaper IFF.

Both Chambers and Chakwas return in ME3 if they survive the final mission of ME2, and Chambers even has an "implicit loyalty" in ME2 that factors into her fate in ME3. So, why are they not explicitly considered?

In Dr. Chakwas's case, her survival is entirely dependent on whether an escort is selected, so rescuing the crew means, at the very least, that she survives.

However, the decisions leading to YN Chambers's loyalty and survival have no bearing on any of the choices the player can make during the finale of ME2. Any decision path can be arbitrarily annotated with all possible combinations of Kelly's loyalty and survival without changing anything else about the choices made in that decision path, and you would always end up with a valid decision path.

In the end, although the previously mentioned choices have consequences in ME3, the author considers them orthogonal to the decisions addressed in this crate.

References

This project was made possible by some amazing people in the Mass Effect community, particularly those who took the time to answer some esoteric questions on the subreddit. Of particular importance was this flowchart for which I regrettably am unable to find any attribution, though I believe that particular version was distilled from a primary source that is also unknown to me.

History

This crate is based on a nearly identical project I originally wrote in Python. The statistics included in that project's README are noticeably different than those above. I am not entirely sure why, though I suspect a subtle bug in the logic that allows the outcome data generation to be paused/continued. That mechanism was necessary to prevent unrelated problems (e.g., power outages) from causing days of computing time to be lost.

Performance

You might be asking, "Days? Really?" Yes, really.

In retrospect, I made some poor decisions in the Python implementation that I avoided this time around:

• I insisted on bit-packing (manually encoding) everything into Python ints, which have a variable bit width.
  • Although the variable size is nice for performing computations with extremely large numbers, it makes bitwise arithmetic quite slow.
  • I didn't actually need to be so paranoid about the size of serialized data structures, so encoding/decoding stuff just wasted time—and made it really annoying to actually query the data.
• The generation logic serialized whatever progress it had made to disk every five seconds.
  • If you had lost data as often as I had, maybe you would have I/O bound your algorithm, too. :'(
• All decision path traversals occurred on the same thread.

The last point turned out not to be as much of an issue in Rust. A naïve port of the Python implementation—minus the bit-packing and overly-eager disk I/O—took less than 90 minutes to generate all of the outcome data on my system. Not bad!

Things got really fast when I finally got around to using more of the CPU's cores. With a combination of the rayon and dashmap crates, it was very easy to parallelize the algorithm. However, applying rayon to everything eventually reaches a point of diminishing returns, so I only changed enough to maximize the speed-up. It turns out that parallelizing the first three levels of recursion was sufficient. As a result, it now takes just over ten minutes to generate all of the data. Good enough!