Released versions of Bear

April XX, 2024: Version 0.5 (XXX KB; MD5: )

“Bear Forest”: A complete overhaul of Bear, borrowing liberally from the philosophy underlying random forests.

March 15, 2024: Version 0.4.9 (365 KB; MD5: 7695088fada346176d6e8a518f52eb29)

Fixed a bug that caused Bear to give up in bootstrapping before finding its first nontrivial model.

October 19, 2023: Version 0.4.8 (365 KB; MD5: 894ba4844ddb40fdda0a16007803d2d7)

Improved the completeness modeling and the messaging when the first run exhausts its time budget before minimizing the loss.

October 11, 2023: Version 0.4.7 (364 KB; MD5: d67ebafcf19e91ba5401833db9610055)

“Python Bear”: Added compare_data to confirm that bear_predict.py gives the same predictions as bear_predict.

October 9, 2023: Version 0.4.6 (363 KB; MD5: f5db72e2e8ba163a96e35ca963fc091a)

“Bear Cull”: removes features not used in the final model and does a second run with just the useful features.

October 4, 2023: Version 0.4.5 (420 KB; MD5: 5daf00e9259044280f5ef7db216f3c4b)

Added support for the MAE (mean absolute error) loss function.

October 1, 2023: Version 0.4.4 (413 KB; MD5: c5d68ee308f3a1e23d15db446e9f74ea)

“Auto Bear”: an exponentially increasing time budget attempts to find the minimum-loss model.

September 25, 2023: Version 0.4.3 (410 KB; MD5: c2d947eac1a1e245c80b485c34f75deb)

Bug fixes.

September 24, 2023: Version 0.4.2 (410 KB; MD5: 426c4923b6d4b5941c1f80459272b2f4)

Bug fixes and proper handling of the balanced log loss function.

August 21, 2023: Version 0.4.1 (400 KB; MD5: 7c441075e04ff3b21ccaf8a7dcaf57d2)

Bug fixes.

August 18, 2023: Version 0.4 (399 KB; MD5: 4bbc10283179f6121d953f6719004197)

“Bear Bones”: returned the core philosophy of Bear back to its 2012 world-beating simplicity.

May 31, 2023: Version 0.3.5 (413 KB; MD5: 07cf73e73f37c9d3296a221110a4cfc0)

Improved the model-building process and removed my initial attempt at Convolutional Bear.

May 9, 2023: Version 0.3.4 (422 KB; MD5: 55fba9d8a28d57c1544bd883d567f516)

Hotfix for mistakes for features with missing values when there is more than one feature.

May 6, 2023: Version 0.3.3 (419 KB; MD5: 538d554d2d6694912dd8e83981830ce6)

Fixed two minor bugs that caused crashes in two of the trivial tutorial examples. Progress reporting improved.

May 5, 2023: Version 0.3.2 (419 KB; MD5: e061944a5d65910c34d05188a9d1f8a6)

Improved the model-growing algorithm.

May 3, 2023: Version 0.3.1 (418 KB; MD5: 622448acd3496e4e8213ea8e709e5f8f)

Performance improvements, including the introduction of the paw 16-bit floating point format for Bear’s internal use.

April 28, 2023: Version 0.3 (433 KB; MD5: 2d08fd4bcf6cb35becc740e80840cb58)

Training and streamed deployment can be separated with memory_bear and bear_predict.

April 15, 2023: Version 0.2.1 (422 KB; MD5: aed69d2596dba97b15e7b410f378546ff)

Performance improvements.

April 9, 2023: Version 0.2 (424 KB; MD5: 16780167c4d76927121c906579d1a210)

Updated to automatically swap memory out to storage as required. Input data should still fit in memory.

March 24, 2023: Version 0.1.1 (409 KB; MD5: 1fc63af9abb4d8753bfe5fa0c4a9e8bf)

Updated to build cleanly on Ubuntu. Instructions included for building on AWS EC2 Ubuntu instances.

March 21, 2023: Version 0.1 (408 KB; MD5: 8e6e5a86ae51ea824e360a84da195f0f)

Initial release. Fully in-memory beta implementation.

© 2023–2024 John Costella