Major efficiency improvements, including

• prune singletons idea
• adding implication finder
• remove backtracking as it is subsumed by other checks and is expensive
• remove subsumed programs from combiner
• improved neg example testing
• improve recalls and pointless relations
• improved unsat core
• add variant pruning
• many tweaks to reduce memory consumption

• temporary horrible hack for annoying janus memory issue
• minor improvements to ALAN ASP encoding
• backtrack prune consistent programs

• filter programs we add to the combine stage
• only build variants using existing variables when the size is bound
• replace internal representation of example coverage from sets to bit arrays

• Slightly improved constraints
• Fixed weird Clingo issue

• Try to deduce BK constraints by default
• Improved testing via Janus
• WIP on improving the constrain stage (only working for single-rule programs)

• Replace pyswip with janus-swi
• Singleton variables enabled by default
• Fixed many silly bugs

• The ability to use different MaxSAT solvers
• Automatically try to detect Datalog BK
• Optimisation for the combine stage when learning programs from noiseless data

Lots of changes, including:

• Support for learning optimal (MDL) programs from noisy data (https://arxiv.org/pdf/2308.09393.pdf) with the --noisy flag.
• Use SAT and MaxSAT solvers to improve efficiency

Major changes:

• Popper now finds minimal unsatisfiable cores of totally incomplete programs, as described in the paper: Learning Logic Programs by Finding Minimal Unsatisfiable Cores, A. Cropper and C. Hocquette, AAAI 2023, which should reduce learning times a lot

Minor changes:

• Faster way to test single rule programs
• Experimental feature to discover constraints from the BK before searching as described in the paper: Learning Logic Programs By Discovering Where Not to Search, A. Cropper and C. Hocquette, AAAI 2023. Enable this feature with the --bkcons flag

A new version of Popper

• The main change is a more efficient way of learning programs with multiple rules, especially non-recursive programs as described Learning programs by combining programs, https://arxiv.org/abs/2206.01614 and which is related to the paper Learning logic programs through divide, constrain, and conquer. A. Cropper. AAAI 2022.