New blog about Causality

I received the following announcement for a new blog on Causality out of the University of California Los Angeles (UCLA) Cognitive Systems Laboratory. It looks very interesting although it might be targeted towards the more academic readers of smart-machines.

The following message was posted on the UAI mailing list,

The UCLA Cognitive Systems Laboratory, in partnership with the Medical Imaging Informatics Group at the University of California, Los Angeles,invites you to participate in a new forum to discuss causality, its principles, recent results, applications, and related philosophical controversies. We feel that developments in this area, over the past two decades, carry the potential of revolutionizing many disciplines in the empirical sciences but, because they stand contrary to established traditions, have not yet captured the imagination of practitioners and educators.

We have constructed a blog devoted specifically to the topic of causation to be driven by visitors with interest in the field. We welcome participants from all backgrounds and views to post questions, opinions, or results for other visitors to chew on and respond to.

Specific topics of interest include:

- Questions regarding the basic principles of causal analysis including its meaning and historical development.
- Views on the controversial status of causation (if any).
- Reviews of current books and papers related to causal inference and its application.
- Discussion and comparison of various approaches and representations.
- Development of practical applications in economics, social science, health sciences, political science, law, and other disciplines based on understanding of cause-effect relationships.
- Typical postings may be found at http://bayes.cs.ucla.edu/BOOK-2K/discussion.html.

Our intent is to promote the advancement of and collaboration within the field by providing a shared forum for interested parties to freely discuss all things related to causality. Being a new effort, we welcome your comments and suggestions on how we can improve this service.

To participate, please visit http://www.mii.ucla.edu/causality/.

We appreciate your interest and look forward to hearing from you!

Alex Bui, PhD
William Hsu
Medical Imaging Informatics Group
University of California, Los Angeles
http://www.mii.ucla.edu

Judea Pearl, Director
Ilya Shpitser
Cognitive Systems Laboratory
University of California, Los Angeles
http://bayes.cs.ucla.edu

WowWee’s marketing genius

WowWee never seizes to amaze me these days. First, they come out of nowhere to market Robosapien, a nifty little humanoid robot that became the hit toy a couple of holiday seasons ago. After that there was no stopping WowWee. They introduced a number of upgrades to Robosapien with version 2 and RS-media. Then they added more robots to their lineup with robopets such as the Roboraptor. Not long ago during the Consumer Electronics Show (CES) in Chicago, the WowWee train kept going announcing and demonstrating a number of new robots including the Roboquad and Robopanda.

As if that is not enough, in what I would call a marketing grand slam, WowWee struck a deal with McDonalds to supply mini versions of their robots for the fast food chain’s Happy Meals. Wowwee created a number of variations of their Robosapien, Roboraptor, Roboreptile and Robopet line of robots specifically for McDonalds. Kids will be able to choose among different minature version’s of the robots including a walking Robosapine, a talking Robosapien, a roaring Roboraptor, a walking robopet and a chomping Roboraptor.

This is clearly a very smart marketing move for WowWee. McDonalds is the largest fast food chain of restaurants in the world and they have a huge appeal towards children. In fact, much of McDonald’s marketing targets children in an effort to lock them in as future customers. The same kids will now grow up having fond memories of WowWee’s robots creating a whole new set of future customers for the company’s robots.

The toys will be available in U.S. and Canada initially and should be out in stores already.

Frances Allen becomes the firs woman to win the Turing Award

IBM’s long time technology guru Frances Allen became the first woman to receive the highly prestigious Turing Award. The 74-year old Allen retired from IBM in 2002 but spent most of her life working on making computers easier to program.

After completing a M.Sc. degree at the University of Michigan, Allen who worked as a math teacher prior to her graduate studies, was hired to work at IBM porting the newly developed Fortran programming language to different computing platforms. During her 45-year career at IBM, Allen also worked on national security projects while her work on parallel computing is well recognized including her contributions to the design of IBM’s Blue Gene supercomputer.

The Turing Award is named after British mathematician Alan Mathison Turing and it has been awarded by the Association for Computing Machinery (ACM) since 1966. The award is currently sponsored by Intel and it comes with a $100,000 prize although receiving the award is far more valuable than that.

Congratulations Francis and thank you for your excellent work!

NVIDIA makes it easy to do math on their GPUs

NVIDIA, maker of computer graphics hardware for 3D rendering, released today a set of libraries and compiler that allows programmers to directly develop applications that do number crunching on the GPU. The new libraries and C compiler are part of the Compute Unified Device Architecture (CUDA) that NVIDIA is promoting as an alternative to ATI’s stream computing initiative for the use of GPUs for general scientific computation.

Providing support for the C programming language, researchers can directly program the GPU and use it as a math coprocessor. CUDA includes numerical libraries for computing the Fast Fourier Transform (FFT) and Basic Linear Algebra Subroutines (BLAS) for matrix multiplication, transposition and vector dot product computation among many others. GPU offers several advantages over traditional CPUs.

Where previous generation GPUs were based on “streaming shader programs”, CUDA programmers use ‘C’ to create programs called kernels that use many threads to operate on large quantities of data in parallel. In contrast to multi-core CPUs, where only a few threads execute at the same time, NVIDIA GPUs featuring CUDA technology process thousands of threads simultaneously enabling high computational throughput across large amounts of data.

It is interesting to see that rivals NVIDIA and ATI are both creating the tools and provide support such that their wonderful hardware can be easily used as a math coprocessor for highly demanding scientific computating. I remember it was no more than 8 years ago when people were starting to exploit graphics accelerators to compute general functions by abusing the OpenGL API. If you wanted to do something really interesting those days you would need access to an SGI graphics workstation or higher. Today you can get similar if not better performance with consumer level graphics cards and the interfaces for programming them are far better especially with the introduction of such kits as CUDA.

Stanford unveils their vehicle, Junior, for DARPA's Urban Challenge

The Stanford team that will be competing in the DARPA Urban Challenge has unveiled their entry vehicle, a 2006 Passat wagon that has been modified by VW engineers such that it can be controlled by a computer. The car is nicknamed Junior in honor of Stanford’s Stanley vehicle that came first in the Grand Challenge but also in tribute to the University’s namesake, Leland Stanford, Jr. The car will be controlled by a network of computers supplied by Intel and using the latest microprocessor multi-core technology. The Stanford team is most concerned with developing the sensing system and artificial intelligence software such that Junior can quickly decide what to do given its situation.

The Urban Challenge is far more difficult than the Grand Challenge because the vehicles will have to navigate in an environment populated by other mobile agents, i.e., moving cars. It is well known in the AI community that solving the planning problem for more than one agents is exponentially more difficult than for single agent environments mostly because the agent has to be able to reason about the actions of the others.

The Stanford team is trying to make things easier for Junior by equipping it with a number of high precision laser sensors all around. This should make it possible for Junior to very accurately detect and avoid collisions with other vehicles. A Point Grey Research Ladybug 2 spherical vision camera mounted on the car’s roof will provide visual feedback necessary for the detection of traffic signs and lights.

Professor Sebastian Thrun who once again is leading the Stanford team has recently said that he envisions robot controlled cars as becoming mainstream in the next couple of decades. He says that “By 2030, roughly, we should be able to deploy this technology on highways, where we would improve human reliability by orders of magnitude.” He says that robot controlled cars will make driving safer and better utilize an already highly congested highway system.



More photos of Junior at News.com, "Under the hood of Stanford's robotic race car"

D-Wave Orion quantum computer: the day after the demo

D-Wave Systems Inc. demoed their quantum computer, codenamed Orion, on February 13th and 15th as they had originally announced. The company showed the 16-qubit computer successfully solving a number of combinatorial problems. D-Wave demoed their quantum computer by remotely controlling it via a terminal application since the machine itself cannot be easily transported and it is currently housed in the company’s facility in Burnaby, British Columbia, Canada. After the demo, many have questioned how quantum this computer is and whether some of D-Wave’s claims of scalability hold merit.

First, the scientific community is very skeptical of D-Wave’s claims that they have constructed the first ever quantum computer. The reason for this is mainly the fact that D-Wave has not published their findings in a scientific publication and as such they have not placed their creation under scientific scrutiny. Their computer is based on the theory of Adiabatic Computation which has been published but scientists already doubt that this approach is the golden key to quantum computation. Even D-Wave’s founders are now back-pedaling from the original claims about their invention saying that it is not certain how quantum their quantum computer really is. In fact,

“D-Wave Chief Executive Herb Martin emphasized that the machine is not a true quantum computer and is instead a kind of special-purpose machine that uses some quantum mechanics to solve problems.”

In addition, as part of the demonstration, D-Wave presented a road-map for their quantum computer promising one thousand qubits and a commercially available product by 2008. Scientists doubt that Adiabatic Computing can scale up to these specs.

Even though it is not clear whether D-Wave’s Orion is the breakthrough for quantum computers that the company claims, it is still rather remarkable that they have managed to go this far as a privately held company. And even if their computer is not a truly quantum machine but only in part, it is still a step forward in the right direction. But don’t keep your hopes up about being able to purchase one of these any time soon. Orion is not a general purpose computer and its applications if any will be similar to today’s supercomputers that are used for scientific, military and large business number crunching.

You can read more about the world’s reaction to D-Wave’s Orion quantum computer demonstration and its implications in the following articles,

CNN: Scientists dubious of quantum computer claims
DailyTeck: Scientists Express Skepticism Over Quantum Computer
Physorg: D-Wave Demonstrated World's First Commercial Quantum Computer
HPCwire: Quantum Computing Steps Out of the Research Lab
Inside Bay Area: Scientists dubious of quantum claims

Faster computer chips spell good news for artificial intelligence research

The last few days, computer juggernauts Intel and IBM have introduced new hardware that promises vast performance improvements in computing speeds.

Intel unveiled their 80-core Teraflop Research Chip (Polaris) that is capable of over one trillion calculations per second (1 Teraflop.) This kind of performance in the past was only achievable with a 2,500 square foot super computer. Intel’s Polaris is not only very powerful but also consumes very little power, the equivalent of today’s 2-core CPUs. Intel executives speculated that the new chip will be available for desktop and laptop PC use in 5 years although the most conservative ones predict that what is most likely to be available in this time frame is a smaller chip with maybe 20 to 40 cores.

In the meantime, IBM showed data about their embedded DRAM (eDRAM) chips that promise to improve computer performance by increasing the amount of low latency memory available to CPUs. The new memory chips are slated to replace SRAM, which is typically used for on-die CPU cache. The new chips occupy one third the space and consume one fifth the power of the currently used SRAM technology. The eDRAM chips should be available from IBM in 2008 along with the new generation of 45nm high performance microprocessors.

These advances in chip design are good news for artificial intelligence research. As I write this, I am waiting for my single core Pentium 4 computer to solve a rather small robot planning problem. Solving this problem takes usually more than 5 hours. The algorithm I use is in fact easily parallelized and if I had access to a CPU with 80 cores, I could definitely get the same problem solved at best 80 times faster. I am definitely looking forward to the coming of these new multi-core CPUs.

iRobot updates its remote controlled Packbot line of robots

There are many uses for robots ranging from entertainment to military and iRobot is one American company that develops robots for much of this spectrum. Other than their toy robots such as My Real Baby and household helper robots such as Roomba and Scooba, iRobot has large contracts with the U.S. military to develop remote controlled robots for use in the battlefield. One of those is the Packbot specifically designed for reconnaissance and explosive ordnance disposal (EOD) missions.

Last week, iRobot introduced the second generation of their military robot the Packbot 510. According to the press release (read here),

PackBot 510 uses a new game-style hand controller for faster training and easier operation in the field. In addition, the robot is 30 percent faster, drags larger objects, lifts twice the weight and has a grip that is three times stronger than its predecessor.

There are at least 800 Packbots in action today in war zones in the Middle East such as Iraq and Afganistan. The robots are saving soldier’s lives and the new robot will help them perform their duties more safely and efficiently. The Packbot 510 will start shipping in April 2007.

Ugobe delays Pleo again

If you were hoping to get your hands on a Pleo robot this coming March as it was previously announced as the month that the loveable Pleo would hit the market then you are out of luck. With a post on Ugobe’s website, Pleo architect Caleb Chung announced that Pleo’s market release will be delayed further to the summer of 2007.

Caleb says that the reason for the delay is Ugobe’s efforts to upgrade Pleo according to their interactions with customers. Ugobe will is working to enhance Pleo’s audio capabilities and touch sensors especially under the chin. Pleo is also receiving a number of cosmetic updates including higher resolution and detail in the eyes and adding soft muscle-like tissue in several key places on its body.

I have to say that I am disappointed with this further delay but the fact Ugobe is listening to its customers and fine tuning its product is not necessarily a bad thing. At the end of the day, who wants to spend $300 and end up with a boring and uninspiring toy?

D-Wave promises to demo the first ever quantum computer

D-Wave, a Canadian and privately held quantum computing company, has announced that on February 13th, 2007, they will perform a live demonstration of the first ever quantum computer. The first demonstration will take place in the Computer History Museum, in Mountain View, California. The company also plans to repeat the demo on February 15th at the Science World at TELUS World of Science in Vancouver, British Columbia.

D-Wave will showcase a 16-qubit quantum processor (the Orion quantum computing system) capable of solving constraint optimization problems that are NP-Complete and can be cast as a two-dimensional Ising model in a magnetic field problem (also known as quadratic integer programming.)

According to Wikipedia,

A quantum computer is any device for computation that makes direct use of distinctively quantum mechanical phenomena, such as superposition and entanglement, to perform operations on data. In a classical (or conventional) computer, the amount of data is measured by bits; in a quantum computer, the data is measured by qubits. The basic principle of quantum computation is that the quantum properties of particles can be used to represent and structure data, and that quantum mechanisms can be devised and built to perform operations with these data.

D-Wave is exploiting the "adiabatic" model of quantum computing to reach this milestone. Adiabatic quantum computation was recently proposed as a general approach to solving NP-hard combinatorial minimization problems (more details about how it works here but you might need a technical background to understand it.)

Quantum computers promise to be asymptotically faster than conventional computers making it possible to solve very hard problems. In fact, a quantum computer with hundreds of qubits should be able to break any of the cryptographic systems in use today.

Scientists predicted that it would be at least another 20 years before a quantum computer could be constructed. D-Wave’s announcement has generated a bit of buzz in the scientific community which is looking forward to the demonstration but remains skeptical about D-Wave’s claims. The live demo is open to the public for both locations as long as one registers online here.

You can find out more about D-Wave’s quantum computer at the company’s official blog.

Robotics and Neuroscience give hope to amputees for a better life

In what has been widely considered a medical breakthrough in rehabilitation for amputees, a woman fitted with a robotic arm has regained the feeling of touch from her missing limb. Specifically, Claudia Mitchell, 26, who lost her left arm in a motorcycle accident, has learned to use her prosthetic robotic arm to perform everyday tasks such as cutting food.

Most importantly, doctors have used a new procedure to reroute the ends of arm nerves to a patch of skin on her chest allowing her to regain the feeling of touch for her missing hand. Doctors hope that when her robotic arm is outfitted with artificial sensors that function similar to the missing arm nerves and connect to the chest tissue, Claudia and other amputees will regain the feeling of touch for their missing limbs. The new technique the doctors used is called targeted muscle reinnervation (TMR).

Dr Leigh Hochberg, a neurologist at Massachusetts General Hospital, said

The next stage would be for touch sensors on the artificial hand transmitting signals back to the re-routed nerves, allowing patients to have accurate sensations of touch, temperature and joint position.

Motorised hooks, hands, wrists and elbows are currently available but movement is usually slow and awkward. Scientists have long been working to create a limb that is controlled by the brain and works well while looking near-normal.

Artificial Intelligence software predicting the outcome of sports events

Recently, I came across an article on The Register discussing the simulated fight in the movie Rocky Balboa. Simulation was used to figure out if the current heavy weight champion Mason Dixon could defeat former champion Rocky Balboa. The outcome of the simulation was a win for Rocky via KO. When the two fighters finally meet at the end of the movie, Dixon wins countering the computer’s prediction although I have to admit that the simulation was about a young Balboa fighting and not the arthritis plagued 60-year old Rocky that entered the ring for his last fight.

Even though the events I just described are fictional, it got me thinking about the potential for intelligent software to predict the outcome of sporting events. I know that people who like to bet on sports would love to have software that can always make the correct predictions.

Although it is next to impossible to correctly model all the variables in a sporting event, current statistical approaches widely used in artificial intelligence can be utilized to analyze past results in order to make educated predictions about future events. Predicting the outcome of a boxing event choosing at random one has a 50% chance of guessing correctly. If an intelligent software agent considers the past history of the fighters including any previous meetings between the two, then it can improve its chances of guessing correctly. However, it can never be 100% certain because no matter how much past data are available, it could never account for all possible variables that can be relevant to the outcome. For example, one of the fighters might become sick the day before the fight and can’t get out of fighting because of pressure from his sponsors.

That said, researchers have used intelligent software to predict the outcome of real sporting events such as the winner of the Super Bowl and the FIFA World Cup. Last summer, Imran Fanaswala and Yashar Fasihnia from the American University in Sharjah collected data going back 20 years and using their FIFA Intelligence (FIFI) software, they predicted that Brazil and Italy would meet in the final of the 2006 FIFA World Cup with Brazil prevailing in the end. Their prediction was wrong since France and Italy were the last two teams with Italy taking the Cup home. Even though their final prediction was incorrect, they did manage to guess one of the two teams in the final game correctly.

Most recently, Electronics Arts (EA) run a software simulation of last Sunday’s Super Bowl game between the Colts and the Bears. They predicted that the Colts would win and they were correct. Although I bet this was more of a publicity stunt for EA than a demonstration of the state-of-the-art in artificial intelligence, it shows that educated predictions of sporting events based on statistical data are possible but not bullet proof. I say this because the same software failed to predict the winner of the Super Bowl last year.

In the future, as we develop algorithms that can discover patterns in large amounts of historical data then we will be able to improve our ability to correctly predict the outcome of sporting events. However, one cannot expect that any future artificial intelligence program will be able to correctly predict the outcome of such events every time.

Report: Robotics for Society conference, day 2

The second day of the Robotics for Society conference organized by the Vancouver Society for Cognitive Systems was as exciting as the first day.

The first speaker was Prof. Alan Mackworth from the University of British Columbia and currently president elect of AAAI. Mackworth started his talk with Asimov’s three laws of robotics and the issue of how to actually design robots that obey them. He then presented his work on the Constraint Based Agent (CBA) framework that provides a language for specifying constraints such as the three laws as well as a mechanism for selecting agent actions such that these constraints are always satisfied. Mackworth admitted that there is still much work that needs to be done before we have the tools to create such CBA agents.

The second speaker was Prof. Hideki Kozima from NiCT, Japan. Kozima presented his work on 2 robots the Infanoid and Keepon. The robots are designed to be social and used for therapeutic and pedagogical applications. Kozima’s group has performed a number of field studies with groups of children having developmental disorders including autism.

The second half of the day included two very interesting talks by professors Stefan Schaal from USC and Masaaki Honda from Waseda University, Japan. Schaal presented much of his work on computational motor control for humanoid robots including how they use reinforcement learning to teach the robots how to complete difficult tasks such as balancing a pole and playing tennis. Honda presented the 6th generation of their state-of-the-art talking head capable of mimicking human speech production.

Richard Rosenberg from UBC closed the day with a short talk about robot ethics. His talk was mostly designed to get people thinking about robot ethics rather than providing answers for any of the important questions such as who is responsible if a robot kills a human being. He said that it was fortunate that even though complex intelligent robots are not likely to be available for many decades people are already talking about the implications of their existence. He compared this with the rise of the computer and all the issues about privacy and security that have become commonplace today but people did not consider until after computers became popular and maybe, by then, it was a bit too late.

Report: Robotics for Society conference, day 1

The Vancouver Society for Cognitive Systems is hosting a two-day conference focused on the interactions between robotics and the societies in which they are embedded. Some very well known robotics and artificial intelligence scientists are in Vancouver to present their perspective on how robots will become part of our society.

The event was kick-started on February 1st with Prof. James Little from the University of British Columbia talking about his efforts to create visually guided mobile robots. Little and his students are designing robots that map and navigate their environment using only stereo vision. They are currently hard at work teaching robots to extract semantic information about their operating environment from dense and accurate geometric information such as 2D occupancy grids and 3D visual landmark maps.

The second speaker was Prof. Gordon Cheng from the ATR Computational Neuroscience laboratory in Japan. Cheng talked about their efforts to construct humanoid robots inspired by biological systems. In their work, they build human-like machines with human-like behaviors that can help them understand human behavior while presenting engineers with a challenging problem, i.e., the construction of the humanoid robots.

After lunch, Prof. Richard Vaughan from Simon Fraser University in British Columbia, Canada, talked about his work of using aggression as a useful metric for increasing a robot’s value. He showed mobile robots negotiating their passing via a narrow passage by modulating their level of aggressiveness to force other robots to backup and clear the way. He concluded his talk with a survey of current military robots and posed some ethical questions about a scientist’s choice to work on developing such systems.

The last talk of the day was given by the well known Rodney Brooks from MIT. Brooks gave a brief overview of the history of robotics wanting to emphasize that just as robots from 50 years ago look primitive compared to robots today, the robots 50 years from now will probably be much more advanced compared to those that are state of the art today. He then continued to discuss why robots are important presenting statistics that clearly show that by 2050 a mostly aged population will require care that could not be provided due to labor shortages. Robots are the obvious answer to this problem. He set some targets for what he believes are required capabilities for such helper robots. These capabilities include basic object recognition at the level of a two year old child, the language capabilities of a six year old child and the social understanding of an eight year old child.

The second day of the conference includes the following speakers, Alan Mackworth, Hideki Kozima, Stefan Schaal, Masaaki Honda and Richard Rosenberg.

What is Artificial Intelligence?

In my interactions with people, I have often discovered that most of them have little to no understanding of what artificial intelligence is. Moreover people know little about what AI’s ultimate goal is. Defining artificial intelligence is rather hard as one can tell from the many definitions available online. In this article I will try to explain what AI is and what makes it difficult to define precisely.

Historically, the term artificial intelligence was coined by John McCarthy in 1956 during the seminal Dartmouth Conference that is widely accepted as AI’s birthplace. McCarthy maintains a website at Stanford that defines AI as,

It (AI) is the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable.

This definition is all encompassing of what AI is. First of all AI is a science. Its scientific goal is to understand the principles that make intelligent behavior possible. Second, AI is an engineering discipline. The central engineering goal is to specify methods for the design of useful, intelligent artifacts (Poole et al.)

In other words, we are not only interested in understanding artificial intelligence systems but we also want to engineer such systems. Engineered systems include physical agents such as robots but also include software agents. McCarthy’s definition also includes the study of human intelligence via the construction of artificial intelligence systems. The word “artificial” is in fact key in the definition because AI is about constructing and studying man-made intelligence that can aid in our understanding of human intelligence. AI is not necessarily about replicating human intelligence. In fact, to this day, we do not have a good understanding of what constitutes “intelligence” let alone “human intelligence.”

The difference between artificial and real intelligence is further discussed in Poole et al,

Is artificial intelligence real intelligence? Perhaps not just as an artificial pearl is a fake pearl, not a real pearl. “Synthetic” intelligence might be a better name, since, after all, a synthetic pearl may not be a natural pearl but it is a real pearl.

So how would you identify an artificial intelligence agent if you saw one or better yet do such agents already exist? To answer these questions we must be clear about what an artificial intelligence agent is given our definition of AI. An AI agent is an agent that acts intelligently, i.e., it makes rational decisions considering its circumstances. That is, an AI agent observes its environment and chooses its actions accordingly. For example, an intelligent agent that is hungry decides to cook instead of doing laundry.

Intelligent agents are already pervasive in our lives. They work under the hood to make our lives easier. For example, online search engines utilize intelligent agents that catalogue information making it easy for us to find it. Other agents work under the hood in our cars making driving safer. Large airplanes fly smoothly and safely because of intelligent agents monitoring the system continuously correcting the flight path. AI agents monitor our email and prevent spam from reaching our Inbox while others make suggestions of what movies we might enjoy watching according to our preferences.

Finally, even though, physical agents such as Robocop still only exist in the movies, there are many intelligent robots performing complex tasks that humans can’t possibly do. For example, NASA’s twin Mars rovers are continuously exploring a distant planet while smaller robots are used here on Earth to inspect oil pipelines and explore hard to reach regions deep underwater.

The future of AI is bright and even though the field has advanced by leaps and bounds during the last 50 years there is still a lot of work left to do. It is hard to predict what artificial intelligence will achieve in the next 50 years but I hope it will come closer to meeting its goals of constructing artificial intelligence agents while helping us understand the nature of our own intelligence.

Resources:

John McCarthy: What is artificial intelligence?
D. Poole, A. Mackworth, R. Goebel (book): Computational Intelligence, a logical approach.