3 releases

0.1.2 Nov 10, 2024
0.1.1 Oct 30, 2024
0.1.0 Oct 30, 2024

#135 in Machine learning

Download history 223/week @ 2024-10-27 41/week @ 2024-11-03 113/week @ 2024-11-10

377 downloads per month

MIT/Apache

425KB
7K SLoC

arcagi

A Rust attempt at the ARC-AGI prize 2024

WARNING:

This code is messy and imcomplete and is likely to change rapidly. At the moment it is a set of experiments to guage what works well and what does not. It is perhaps 20% complete and will answer just over 10% of the examples given.

Background

The Kaggle ARC-AGI challenge 2024 is an attempt to get computers to solve visual 2D puzzles, similar to those in many IQ tests. To solve this challenge in a principled way requires perception of stucture and reasoning guided by constraints. It currently cannot be solved by LLM/GenAI based methods though there may be a small place for them (or possibly just simple Neural Networks for specific recognition tasks).

This code is a first submission for the prize, it could conceivably solve all the puzzles in the challenge, but at the moment solves just over 10% in the public evaluation set. Much more work is required to reach the 85% or better level.

This prize in important because it will demonstrate reasoning abilities and not just the memoristion with some 'light' generalisation that is currently demonstrated by all Gen-AI models (in fact all Machine Learning models).

Abstraction, Generalisation and Functional composition are key concepts. These work in tandem with constraint recognition and discovery to tame the size of the search space.

This is a pure Symbolic AI approach and hence is deterministic. There may be a place for an LLM/Gen-AI in object recognition and/or generating candidate solutions. It was hoped that the author's previous experience of LLMs (see LLMClient), would help (for code generation). It might eventually, for now there is plenty to explore with the current Symbolic approach. Code generation was experimented with, given some data structures and a textual description of the problem, but the code generated was poor quality and time consuming to debug (Rust is probably not the best target for code generation given it's memory management as simple next token prediction does not understand that).

See: https://www.kaggle.com/competitions/arc-prize-2024

See inet submission.

To run this the first parameter is one of 'training', 'evaluation' or 'test'. There is no solutions file for the test run, this is expected so ignore the error message. A second parameter can be supplied which is the test to run, useful for testing individual cases. A typical call to run the training set would be 'cargo run --release training'.

Kaggle notebook

from cffi import FFI
ffi = FFI()
ffi.cdef("int test();")
rust = ffi.dlopen("/kaggle/input/release/libarc_agi.so")
a = rust.test();
print(a)

NOTE

To compile the above for kaggle a docker image needs to be created and the .so file uploaded and referenced, somewhat tedious. It will run locally for public data test with cargo as normal. A release version should run in less that a second on most machines, this will change as more cases are completed.

Currently there is a problem with the above, with kaggle scoring on the private dataset. Only run locally on public data for now or join the inet team.

More investigation of the scoring on the secret test set was done, in addition to more training and evaluation examples succeeding. There is still a problem in that the score is always 1.00 regardless of what is done. Currently 47 evaluation tasks are correctly answered. The code being deterministic a binary chop over the top level rules can be done. If this is done over to distinct halves then on both occasions 1.0 is scored, not 0.0 and 1.0 or vice versa as would be expected. Any halves! The submission file is correctly formatted (though lacks optional white space). A null run produces a score of 0.0 as would be expected. So, very confused.

Dependencies

~2.3–3.5MB
~67K SLoC