MAJOR-CHANGES.txt

Jan 21 2016
	- Extreme simplification of proposals. 
Jan 7 2016
	- Added "display" to hypotheses to specify printing. This replaced FunctionHypothesis.args and LOTHypothesis.args, and makes it easier to write less function code, since we can display with arbitrary strings
	- Removed special printing from FunctionNode -- everything should use string interpolation now. Removed: If, apply, lambda, comma, ''
	- Included BV printing -- now a node that introduces a BV can use it in its interpolation string via <BV>
	- Updated version (technically, not backwards compatible)
Dec 26 2015
	- Added examples for Prolog and PureLambdaCalculus, along with a few associated fixes
Dec 16 2015
	- added shortcut sampling scheme
	- reconfigured pystring and classes to allow bv without lambdas
    - Moved some examples and Sandbox to Projects subfolder
    - Removed Evaluation and moved its parts to LOTlib.Eval (that's where @primitive lives now)
    - Removed testing and integrated unit tests throughout
    - Changed FiniteBestSet to TopN; rewritten for clarity
October 8 2015
    - Added stream processing functions to MCMCSummary. SampleStream now allows a convenient notation for processing streams of samples. 
    - Moved standard_sample to Inference.Samplers.StandardSample
    - Moved GrammarInference to its own folder, old grammar inferece code to "Legacy"
    - Added binding to Stan to do MCMC on grammar parameters, MAP inference, and some simple (experimental) MAP bootstrapping
    - Added documentation! A new tutorial!
July 31 2015
    - Changed formatting of examples, etc -- now just a single file for most examples.
    - Added format functionality so that a name can be of the form "f(%s)", where %s is the arguments. This allows for arbitrary syntax (not just function composition)
July 11 2015
    - Removed storage of generating rule from FunctionNode in order to allow save FunctionNodes to be loaded and have prior computed on a new grammar. - Now uses FunctionNode.get_rule_signature() and GrammarRule.get_rule_signature()
	- Changed BVContextManager to not store a list of added rules
	- Documentaiton lives in LOTlib/Documentation. (Needs to be recompiled)
	- Version updated to 0.5.1	
June 14 2015
    - Changed Lexicon default proposal to sampling each word with a fixed prob. 
    - Got rid of evaluate_expression -- now just directly use eval
    - Substantially simplified __copy__ so that it really only needs to happen in Hypothesis. Initializers should allow no args. 
    - Change Lexicon -- now it no longer takes a list of words and a hypothesis construction function. Both should be done externally. 
    - Changed the proposal_function method so now proposals should be done via inheritance. The old proposal schemes are around in case its useful but will be removed in the future. 
    - Alpha was moved to be inside of each FunctionData element rather than in a hypothesis. It is a property of data, not hypotheses. 
    - Extracted a few different likelihood methods to Hypotheses.Likelihoods. 
    - Changed SimpleGenerativeHypothesis to StochasticFunctionLikelihood. 
    - Updated version to 0.4.0
    - Made Search defaultly work with all the Example models
    - Change MPI_map to NOT defaultly randomize the order. The can be done in the calling context. 
May 31 2015
    - Changed compute_posterior to now return the posterior, not p and l. 
    - created a decorator @Hypothesis.attrmem to use instead of setting self.prior etc by hand
    - LOTlib NO LONGER guarantees that calls to compute_prior or compute_likelihood will set posterior_score. Too much baggage. Use compute_posterior or update_posterior explicitly
    - Moved LOTlib/Examples/Tutorial to LOTlib/Tutorial
    - Standardized model running and inference so that everything uses the same form. See
    Examples/Readme.md    
Mar 10 2015
    - NOTE: Changed log_probability to only be run from grammar. FunctionNodes now no longer store generation_probability. This will require changes to use grammar.log_probability(t) instead of t.log_probability()
    - Changed lot_iter to break_ctrlc    
Mar 6 2015
    - added abstract Sampler class to make samplers compositional. Now you can make a PartitionMCMC of a ParallelTempering, for instance
    - Added experiment Monte Carlo Tree Search in Inference.MCTS. Includes a  bunch of minor changes to make this work better. 
Feb 1 2015
    - On failure of a proposal type, proposers now raise ProposalFailedException, rather than returning [newt,0.0]
    - Changes to FunctionNode.__eq__ that now make it take into account "invisible" nodes (FunctionNodes that have a name of '')
    - Reorganization of performance testing
Jan 7 2015
    - Changed @LOTlib_primitive to @primitive
Dec 28 2014    
    - Removed resample_p -- now you must specify a custom function in FunctionNode (resampleProbability). Removed in GrammarRule too. 
    - Changed impossible proposals to raise a ProposalFailedException
    - fixed copying twice in LOTProposal
    - @LOT_primitive now adds to eval, no need to call register_primitive; moved LOTlib.Primitives.Primitives to LOTlib.Evaluation.Eval
    - Moved Number primitives to Examples.Number.Model.Primitives
    - Removed listifnot from MPI. Now MPI_map and MPI_unorderedmap *must* be given an iterable of lists
Dec 18 2014
    - Updates to testing framework -- some MCMC tests, proposal, grammar. Old version is in Testing/old
    - moved lp_regenerate_propose_to_grammar to RegenerationProposal
Dec 12 2014
    - Removed mh_sample (use MHSampler instead)
    - moved FunctionNode->String functions to Visualization.Stringification, as well as FunctionNode->Dot
    - Moved FiniteBestSet to LOTlib.Utilities
    - created MCMCSummary
Nov 19 2014:
    - RecursiveLexicon added to enable mutual recursion. 
    - SimpleLexicon now no longer copies *all* values on proposal--new proposed one may point to old values
Nov 17 2014:
    - Recursion defaultly removed from LOTHypothesis. Now, recursive calls must use RecursiveLOTHypothesis. (This saves all LOTHypotheses from having to wrap in a Y combinator)
    - Added Annealing and Schedule classes
    - Re-organization of Number
Nov 13 2014:
    - fixes to inverse-inline proposal
    - value2function(value) now is compile_function(), and assumes that the value has been set
    - Updates to organization of Examples
    - More examples brought up to date
    - FunctionHypothesis.__call__ now does no error handling--exceptions must be handled by subclasses
    - Fixed NameErrors to be useful!
Oct 31 2014:
    - Integration of SimpleMPI with LOTlib (now LOTlib.MPI). SimpleMPI will no longer be updated. 
Sep 24 2014:
    - Major revision to Grammar.enumeration (no longer called increment_tree)
    - Changed string output to include the "implicit" lambda on LOTHypotheses
Aug 31 2014:
    - Fixed setto to update parents correctly
    - Changed "replace_subnodes" to use a predicate rather than equality, since equality is a bit of a hazard with new bound variables
    - Fixed FunctionNode copy to correctly manage parents
    - fixed lambda thunk print error
    - update InverseInline proposal, but not thoroughly checked
Aug 29 2014:
    - Removed isapply, islambda, isbv, etc. The lambdas and bvs can be checked by checking the type of the FunctionNode or GrammarRule
    - changed BVRuleContextManager to take a functionNode as its first argument, not a rule. It defautly looks at the .added_rule attribute of the FunctionNode
    - Added parent references to FunctionNodes. This simplified some code in resampling proposals (and allowed them to use BVContextManager)
    - Move subnode sampling from grammar to FunctionNode
    - These changes got rid of the need to do crazy "for ... in" notation for sampling nodes. Now it works via BVContextManger as in RegenerationProposal
    - InsertDeleteProposal is now defaultly NOT a mixture of insertdelete and RegenerationProposal
    - FunctionNdoes now store a pointer to the rule that made them. This is required to allow them to recompute their probabilities under a grammar
    - Cleaned up some code in Grammar    
    - Primitives moved to inside Evaluation
    - Proposals moved inside Inference
    - Move tests into Testing, and removed some extras
    - fixed problem in "if_" print order
    - Updated Number Shared code
Aug 22 2014:
    - Fixed bug in enumeration with new bound variable setup; make bound variables increment in depth order 
Aug 19 2014:
    - FunctionNode bound variables are now stored with uuids *internally* rather than a string plus the depth. We use the string+Depth view when casting a FunctionNode to a string.
    - pystring was moved out of FunctionNode to make it a little more elegant
    - FunctionNode.__eq__ the fancy version is now not as necessary (since all bv_names are unique, and that's handled by pystring), so that has been removed. We still could check uniqueness wrt bv_prefix. 
Aug 16 2014:
    - Refactoring of GrammarRule and creation of a BVRuleContextManager to allow easier management of added bound variables. Changes in grammar to use "with" instead of adding and subtracting rules
    - Change to FunctionNode for nicer printing via %s
    - Refactoring of FunctionNode to store the *rule* added, rather than the bv_name, bv_type, etc. Now use FunctionNode.added_rule to get hte bv characteristsics
    - Removed Grammar.add_bv_rule and Grammar.remove_rule
    - Added BV to increment_tree with fancy new "with" statements
    - GrammarRule now will return a FunctionNode "stub" that hasn't expanded the args yet. This is used in generation. 
    - Some early edits for applylambda. Not clear if that will be feasible to add. 
July 31 2014
    - Refactored Proposals to have separate classes. 
July 20 2014
    - Refactored Inference.* to use classes rather than functions. This way, samplers can save states, be resumed, etc. 
June 20 2014
    - Created Evaluation to implement different virtual machines for evaluating FunctionNodes. The most basic if Eval, whcih calls python eval. 
    - Refactored BasicPrimitives to LOTlib.Primitives; added ability to add new primitives
June 15 2014
    - Deleted Memoization
    - Added CombinatoryLogic
    - Modified Parsing
    - Fixed Grammar to delete rules in case an exception gets raised during generation
    - Grammar.BV_P is now the default, but these can be set for individual rules via bv_p in add_rule
    - renamed ll_temperature -> likelihood_temperature in MetropolisHastings, to match with prior
May 17 2014:
    - Removed Serialization -- reverted to old pickling scheme for FunctionHypothesis because otherwise MPI doesn't work well
    - Added simple CCG-ish example
    - Updated SymbolicRegression to use new compute_single_likelihood
    - Moved RationalRules code and helper functions out of Grammar and into RationalRulesLOTHypothesis
    - Fixed prior,likelihood temperature passing in hypothesis initialization
    - fixed number demo to work with compute_single_likelihood
    - fixed MAPSymbolicRegressionHypothesis
May 13 2014:
    - Changed LOTHypothesis.compute_single_likelihood to require a FunctionData
    - Altered how functions are passed to LOTHypothesis and FunctionHypothesis -- compute_single_likelihood now takes a single argument, a FunctionData
    - Fixed RationalRules examples to use new Hypotheses.RationalRulesLOTHypothesis
May 12 2014:
    - Added support for types (in the semantic/functional sense), that play nicely with nested lambdas
    - Major refactoring of Lexicon, into SimpleLexicon, WeightedLexicon, and Examples/Quantifier/GriceanWeightedLexicon
    - Changed SimpleLexicon to hold Hypotheses for each word, instead of FunctionNodes
    - Addition of a Testing folder
    - Added example on regular expression inference, demonstrating how to add a new type of "eval"
    - Moved vectorized quantifier stuff to a new subfolder, that will have to be brought up to date eventually
    - Changed serialization to allow for serialization of arbitrary things, with lambdas, etc. This uses the "cloud" library instead of pickle
May 1 2014:
    - Major change: is_terminal now only checks that the descendants are NOT functionNodes
    - Added lot_iter, which takes a generator and wraps it so that ctrl-c will break 
    - Added magnetism example
Jan 28 2014:
    - Many changes to naming:
    lp now posterior_score
    v (function argument in hypothesses) now value
    tweaks to weighted_sample, to remove old style usage
    chanes to p/lp in GrammarRule
Dec 14 2013:
    - Major revision breaking backward compatibility -- changes to FunctionNode syntax in PCFGs so it makes more sense. Non-function terminals have None as args; function terminals have []
    - Also changed "PCFG" to "Grammar" since it's not technically a pcfg
    - Changed bv so that they now have a name and a type (if the type is set to [], they are function calls, else they can expand their type below)
    - In doing this, the grammar was changed so that each nonterminal can add at most *one* bound variable (no longer a list). This lets lambda be higher order (but typed)
    - Changed to semantic version numbering, and advanced to version 0.0.1
Dec 12, 2013:
    - Changed the .copy() methods to all use the python standard __copy__. This now correctly is handled throughout inheritance
    - Fixed README.md
    - Removed memoization -- not used anymore (but should be added back in eventually)
    - Removed a bunch of old files
Dec 10, 2013:
    - Changed GaussianLOTHypothesis to require ll_sd in FunctionData, not as an argument to the hypothesis.
Dec 9 2013:
    - Removed crazy function call notation () in bound variables. Now if you want that, you need to call apply (as before)
    - Cleaned up FiniteBestSet a little -- added FiniteBestSet.add and kinder merge
    - Cleaned up Number Search to play nicer with new MPI_map
Dec 2 2013:
    - Added RR example
Sep 23, 2013:
    - Removed parallel MPI functions to a separate library, SimpleMPI. This keeps the core LOTlib code separate, and allows easy use of mapping in other applications
July 28, 2013
    - added NAND Compression example
    - Made FunctionNode.depth() return an int
    - added all_binary_vectors to Miscellaneous
April 23, 2013
    - Got rid of the annoying MPI warning for too many calls to MPI_Finalize()
    - Progress bar tweaks
    - made BasicPrimitives.ancestors O(N) instead of O(N^2)  
April 22, 2013
    - Changed "StandardExpression" to "LOTHypothesis"
    - Made FunctionNodes have a default iterator of subnodes, and removed FunctionNode.all_subnodes()
    - Changed "temperature" in MCMC to apply *only* to the LL, not the prior. This helps keep our PCFGs defined even for higher temperature
    - Created a friendly changelog