aima-python

Python code for the book Artificial Intelligence: A Modern Approach. We're loooking for one student sponsored by Google Summer of Code (GSoC) to work on this project; if you want to be that student, make some good contributions here (by looking throush the "Issues" and resolving some), and submit an application. And we're always looking for solid contributors who are not affiliated with GSoC. A big thank you to everyone who has contributed!

Python 3.5

This code is in Pythoin 3.5. If you don't have that version, you should install it, and if you can't install it, use a browser-based Python interpreter such as repl.it.

Structure of the Project

When complete, this project will have Python code for all the pseudocode algorithms in the book. For each major topic, such as logic, we will have the following three files in the main branch:

• logic.py: Implementations of all the pseudocode algorithms, and necessary support functions/classes/data.
• logic.ipynb: A Jupyter notebook, with examples of usage. Does a from logic import * to get the code.
• tests/logic_test.py: A lightweight test suite, using assert statements, designed for use with py.test.

Until we get there, we will support a legacy branch, aima3python2 (for the third edition of the textbook and for Python 2 code). To prepare code for the new master branch, the following two steps should be taken:

Port to Python 3; Pythonic Idioms; py.test

• Check for common problems in porting to Python 3, such as: print is now a function; range and map and other functions no longer produce lists; objects of different types can no longer be compared with <; strings are now Unicode; it would be nice to move % string formating to .format; there is a new next function for generators; integer division now returns a float; we can now use set literals.
• Replace old Lisp-based idioms with proper Python idioms. For example, we have many functions that were taken directly from Common Lisp, such as the every function: every(callable, items) returns true if every element of items is callable. This is good Lisp style, but good Python style would be to use all and a generator expression: all(callable(f) for f in items). Eventually, fix all calls to these legacy Lisp functions and then remove the functions.
• Create a _test.py file, and define functions that use assert to make tests. Remove any old doctest tests. In other words, replace the ">>> 2 + 2 \n 4" in a docstring with "assert 2 + 2 == 4" in filename_test.py.

New and Improved Algorithms

• Implement functions that were in the third edition of the book but were not yet implemented in the code. Check the list of pseudocode algorithms (pdf) to see what's missing.

• As we finish chapters for the new fourth edition, we will share the new pseudocode, and describe what changes are necessary.

• Create a .ipynb notebook, and give examples of how to use the code.

Style Guide

There are a few style rules that are unique to this project:

• The first rule is that the code should correspond directly to the pseudocode in the book. When possible this will be almost one-to-one, just allowing for the syntactic differences between Python and pseudocode, and for different library functions.
• Don't make a function more complicated than the pseudocode in the book, even if the complication would add a nice feature, or give an efficiency gain. Instead, remain faithful to the pseudocode, and if you must, add a new function (not in the book) with the added feature.
• I use functional programming (functions with no side effects) in many cases, but not exclusively (sometimes classes and/or functions with side effects are used). Let the book's pseudocode be the guide.

Beyond the above rules, we use Pep 8, with a few minor exceptions:

• I'm not too worried about an occasional line longer than 79 characters.

• You don't need two spaces after a sentence-ending period.

• Strunk and White is not a good guide for English.

• I prefer more concise docstrings; I don't follow Pep 257.

• Not all constants have to be UPPERCASE.

• Pep 484 type annotations are allowed but not required. If your parameter name is already suggestive of the name of a type, such as url below, then i don't think the type annotation is useful. Return type annotations, such as -> None below, can be very useful.

    def retry(url: Url) -> None:
  

Index of Code

Fig	Page	Name (in book)	Code
2	32	Environment	Environment
2.1	33	Agent	Agent
2.3	34	Table-Driven-Vacuum-Agent	TableDrivenVacuumAgent
2.7	45	Table-Driven-Agent	TableDrivenAgent
2.8	46	Reflex-Vacuum-Agent	ReflexVacuumAgent
2.10	47	Simple-Reflex-Agent	SimpleReflexAgent
2.12	49	Reflex-Agent-With-State	ReflexAgentWithState
3.1	61	Simple-Problem-Solving-Agent	SimpleProblemSolvingAgent
3	62	Problem	Problem
3.2	63	Romania	romania
3	69	Node	Node
3.7	70	Tree-Search	tree_search
3	71	Queue	Queue
3.9	72	Tree-Search	tree_search
3.13	77	Depth-Limited-Search	depth_limited_search
3.14	79	Iterative-Deepening-Search	iterative_deepening_search
3.19	83	Graph-Search	graph_search
4	95	Best-First-Search	best_first_graph_search
4	97	A*-Search	astar_search
4.5	102	Recursive-Best-First-Search	recursive_best_first_search
4.11	112	Hill-Climbing	hill_climbing
4.14	116	Simulated-Annealing	simulated_annealing
4.17	119	Genetic-Algorithm	genetic_algorithm
4.20	126	Online-DFS-Agent	online_dfs_agent
4.23	128	LRTA*-Agent	lrta_star_agent
5	137	CSP	CSP
5.3	142	Backtracking-Search	backtracking_search
5.7	146	AC-3	AC3
5.8	151	Min-Conflicts	min_conflicts
6.3	166	Minimax-Decision	minimax_decision
6.7	170	Alpha-Beta-Search	alphabeta_search
7	195	KB	KB
7.1	196	KB-Agent	KB_Agent
7.7	205	Propositional Logic Sentence	Expr
7.10	209	TT-Entails	tt_entials
7	215	Convert to CNF	to_cnf
7.12	216	PL-Resolution	pl_resolution
7.14	219	PL-FC-Entails?	pl_fc_resolution
7.16	222	DPLL-Satisfiable?	dpll_satisfiable
7.17	223	WalkSAT	WalkSAT
7.19	226	PL-Wumpus-Agent	PLWumpusAgent
9	273	Subst	subst
9.1	278	Unify	unify
9.3	282	FOL-FC-Ask	fol_fc_ask
9.6	288	FOL-BC-Ask	fol_bc_ask
9.14	307	Otter
11.2	380	Airport-problem
11.3	381	Spare-Tire-Problem
11.4	383	Three-Block-Tower
11	390	Partial-Order-Planner
11.11	396	Cake-Problem
11.13	399	Graphplan
11.15	403	SATPlan
12.1	418	Job-Shop-Problem
12.3	421	Job-Shop-Problem-With-Resources
12.6	424	House-Building-Problem
12.10	435	And-Or-Graph-Search	and_or_graph_search
12.22	449	Continuous-POP-Agent
12.23	450	Doubles-tennis
13.1	466	DT-Agent	DTAgent
13	469	Discrete Probability Distribution	DiscreteProbDist
13.4	477	Enumerate-Joint-Ask	enumerate_joint_ask
14.10	509	Elimination-Ask	elimination_ask
14.12	512	Prior-Sample	prior_sample
14.13	513	Rejection-Sampling	rejection_sampling
14.14	515	Likelihood-Weighting	likelihood_weighting
14.15	517	MCMC-Ask
15.4	546	Forward-Backward	forward_backward
15.6	552	Fixed-Lag-Smoothing	fixed_lag_smoothing
15.15	566	Particle-Filtering	particle_filtering
16.8	603	Information-Gathering-Agent
17.4	621	Value-Iteration	value_iteration
17.7	624	Policy-Iteration	policy_iteration
18.5	658	Decision-Tree-Learning	DecisionTreeLearner
18.10	667	AdaBoost	AdaBoost
18.14	672	Decision-List-Learning
19.2	681	Current-Best-Learning
19.3	683	Version-Space-Learning
19.8	696	Minimal-Consistent-Det
19.12	702	FOIL
20.21	742	Perceptron-Learning	PerceptronLearner
20.25	746	Back-Prop-Learning
21.2	768	Passive-ADP-Agent	PassiveADPAgent
21.4	769	Passive-TD-Agent	PassiveTDAgent
21.8	776	Q-Learning-Agent
22.2	796	Naive-Communicating-Agent
22.7	801	Chart-Parse	Chart
23.1	837	Viterbi-Segmentation	viterbi_segment
24.21	892	Align

Choice of Programming Languages

Are we right to concentrate on Java and Python versions of the code? I think so; both languages are popular; Java is fast enough for our purposes, and has reasonable type declarations (but can be verbose); Python is popular and has a very direct mapping to the pseudocode in the book (but lacks type declarations and can be slow). The TIOBE Index says the top five most popular languages are:

    Java, C, C++, C#, Python

So it might be reasonable to also support C++/C# at some point in the future. It might also be reasonable to support a language that combines the terse readability of Python with the type safety and speed of Java; perhaps Go or Julia. And finally, Javascript is the language of the browser; it would be nice to have code that runs in the browser, in Javascript or a variant such as Typescript.

There is also a aima-lisp project; in 1995 when we wrote the first edition of the book, Lisp was the right choice, but today it is less popular.

What languages are instructors recommending for their AI class? To get an approximate idea, I gave the query norvig russell "Modern Approach" along with the names of various languages and looked at the estimated counts of results on various dates. However, I don't have much confidence in these figures...

Language	2004	2005	2007	2010	2016
none	8,080	20,100	75,200	150,000	132,000
java	1,990	4,930	44,200	37,000	50,000
c++	875	1,820	35,300	105,000	35,000
lisp	844	974	30,100	19,000	14,000
prolog	789	2,010	23,200	17,000	16,000
python	785	1,240	18,400	11,000	12,000

Acknowledgements

Many thanks for contributions over the years. I got bug reports, corrected code, and other support from Darius Bacon, Phil Ruggera, Peng Shao, Amit Patil, Ted Nienstedt, Jim Martin, Ben Catanzariti, and others. Now that the project is in Githib, you can see the contributors who are actively improving the project. Thanks!