Jordle

Java wordle solver implementation following along with the "roget" implementation stream (in Rust), which implements the 3blue1brown algorithm (video1, video2). The implementation was optimized a bit along the way.

Variants

Currently, there are two algorithmic variants of the wordle solver: 1) Entropy (3b1b) and 2) MostFreq.

Entropy

Follows the 3blue1brown algorithm to choose guesses based on the expected gained information of each word.

Mostfreq

Simple guesser that chooses the word with the highest frequency in the dictionary.

Scores & Performance

The scores as computed against all answers with a max. of 6 guesses each (which is configurable in the source):

Entropy: avg score: 3.892023, solved: 98.27%
Mostfreq: avg score: 4.277778, solved: 96.67%

'mostfreq' ran 3.19 ± 0.10 times faster than 'entropy'

So, the Entropy strategy implementation is better but takes nearly 400% as long!

Build & Run

mvn --quiet clean package
bin/jordle

Benchmarking

Use hyperfine to compare different implementations.

⌘  hyperfine -w 0 -n entropy 'bin/jordle -a entropy 64' -n mostfreq 'bin/jordle -a mostfreq 64'
Benchmark 1: entropy
  Time (mean ± σ):      1.408 s ±  0.044 s    [User: 1.439 s, System: 0.265 s]
  Range (min … max):    1.357 s …  1.507 s    10 runs

Benchmark 2: mostfreq
  Time (mean ± σ):     441.3 ms ±   4.1 ms    [User: 615.8 ms, System: 106.6 ms]
  Range (min … max):   437.2 ms … 449.8 ms    10 runs

Summary
  'mostfreq' ran
    3.19 ± 0.10 times faster than 'entropy'

TODOs

Other algorithms/implementations from the stream.
Use test cases from wordle-tests
Parallelization of candidate evaluation (not games, which is rather trivial).

GraalVM native-image

I tested the implementation under various different GraalVM versions/JDKs: the "old" JDK under Rosetta and a pre-release for the M1 ARM architecture. Benchmark results comparing the ARM version compiled to an executable and "standard JAR" can be found in the plot.ipynb.

Rosetta (Intel-JDK)

Version: CE 22.0.0.2 (build 17.0.2+8-jvmci-22.0-b05)

Steps to reproduce:

Install GraalVM under homebrew using a Rosetta-enabled Terminal.app (arch should print i386). It seems that this terminal is only used for installing via homebrew, such that it does not complain about a missing build for the arm64 architecture. Apparently, GraalVM 22.1 (or 22.2 latest) will support M1 natively.
Remove quarantine flag: sudo xattr -r -d com.apple.quarantine /path/to/graal.
Set PATH to include the GraalVM binaries in bin and JAVA_HOME to point to GraalVM Home. I recommend doing this temporarily, so you can more easily experiment and do not hose your system.
Make sure to check the current Java version java -version. It should include "GraalVM".
Install native-image: gu install native-image.

Package the project (mvn clean package) and compile to a native image:

native-image -jar target/release/lib/jordle.jar --install-exit-handlers -H:IncludeResources='.*\.txt$' -H:+ReportUnsupportedElementsAtRuntime

The Rosetta native-image build takes around 1m 10s on the M1 MacBook Air.

⌘  file jordle
jordle: Mach-O 64-bit executable x86_64
⌘  du -h jordle
 15M	jordle

NB: "DARWIN does not support building static executable images."

Darwin+aarch64 (M1 Native JDK)

Steps to reproduce below.

Note: I'm using fish instead of bash! Also, I'm setting it up as a custom asdf Java version.

Download current preview and extract:

$ pushd $HOME/.asdf/installs/java
$ curl -sLO https://github.com/graalvm/graalvm-ce-dev-builds/releases/download/22.1.0-dev-20220321_2332/graalvm-ce-java17-darwin-aarch64-dev.tar.gz
$ tar xzvf graalvm-ce-java17-darwin-aarch64-dev.tar.gz

Set up for asdf:

$ cd graalvm-ce-java17-22.1.0-dev
$ for d in (ls Contents/Home); ln -s "Contents/Home/$d" $d; end
$ asdf reshim

Make sure it's listed and go back to the code directory:
```
$ asdf list java
$ popd
```

Configure as local version and test:

$ asdf local java graalvm-ce-java17-22.1.0-dev
$  java -version
openjdk version "17.0.3" 2022-04-19
OpenJDK Runtime Environment GraalVM CE 22.1.0-dev (build 17.0.3+4-jvmci-22.1-b03)
OpenJDK 64-Bit Server VM GraalVM CE 22.1.0-dev (build 17.0.3+4-jvmci-22.1-b03, mixed mode, sharing)

Install native-image:
```
$ gu install native-image
$ asdf reshim
```

Package the project (mvn clean package) and compile to a native image:

$ native-image -jar target/release/lib/jordle.jar --install-exit-handlers -H:IncludeResources='.*\.txt$' -H:+ReportUnsupportedElementsAtRuntime

The native native-image build takes around 21s on the M1 MacBook Air.

⌘  file jordle
jordle: Mach-O 64-bit executable x86_64
⌘  du -h jordle
 15M	jordle

NB: "DARWIN does not support building static executable images."

Flamegraphs

Using async-profiler, we can generate flamegraphs pretty easily. Unfortunately, on macOS it seems we're limited to user-space code only :(. Nonetheless, they provide interesting introspection into the running code.

For example, to evaluate a currently running jordle (from the bin script, which sets some necessary JVM args), execute:

../async-profiler-2.7-macos/profiler.sh -d 10 -t -f perf-out.html (jps | awk -F ' ' '/jordle/ { print $1}')

After the 10 second sample time, you can then open perf-out.html in the browser to inspect your flamegraph.

enplotz / wordle_solver