What do Microsoft need now?

I’d never really thought about the internet advertising market til the Microsoft-Yahoo thing came up. I mean, I know that when I search on G, a load of adverts appear above and beside my search results and there’s an order to the search results that may or may not have a commercial bias (at least it will if the companies are canny about things like keywords and Dmoz), but I’d never really thought about how they got there and what it means.

So. The Microsoft-Yahoo marriage is off, the bride’s run away, the Maid of Honour (News Corp) won’t be forgiven because she took her dress off and played best man instead, and the groom’s still on the hunt for a pretty girl to hook up with (but choose wisely my friend; divorce costs). Except it’s all gone a bit Babylonian bride market.

So again. If Microsoft wants to take on the internet advertising market, what does it need? In one word: eyeballs. In more words: eyeballs, search histories, processing and trust. And that’s the hard part. People stay in business because they understand transactions (or genuinely do have a product that a) people want and b) not many other people can provide. But that’s rare). And there are at least three sets of people in these transactions: the companies paying to gain access to people who might want to view their sites, the viewers and the matchmakers (staying on the marriage theme for a moment) between them.

Google (and this is at base all about Google)’s internet advertising works because it brings in an incredible volume of traffic (through both search and all the other tools it punts, e.g. Google Earth etc), knows what that traffic thinks it wants (search histories and profiles) and doesn’t chuck too many unsuitable matches at the user (does Google have a dating service? Maybe it should…silly me; it’s horrifying). Microsoft wants to match this. Although it could be argued that a more lateral approach than building its own Google would be more sensible, I’ll indulge that idea for a moment. So who would I buy if I were Steve B at the moment?

Well. Potentially Murkysoft has its own eyeballs and processing, but not much in the way of search histories or trust. So a search engine or information site, preferably something large and generic (Ask? Dogpile?) or failing that a set of more topic- specific ones (Kelkoo? Multimap? About?). Trust is a difficult one; to pull this off, Microsoft would either have to get its own teams working closer together (it seems it suffers from the same internal markets that made NASA fail) or front its efforts through a more-trusted third party (in a hand-sweepy way, nobody trusts big business but everyone likes a friendly puppy). It has enough access to ideas (through its own research teams and university links); it just needs the front, the stringy bit that holds them together. Me, I’d be kicking myself for not having started a small esoteric search site, if I hadn’t been doing so many cool things instead. But it will be fun to watch all the other girls applying their makeup and practicing their winks. I just hope they don’t fight too badly if the answer is a harem.

Old papers: Intelligence Analysis for New Forms of Conflict

Another old friend from back in the days before old age and disillusionment (aka 1997). I don’t do this sort of thing any more, but I’d really like to think about how it could be applied to more civilian activities.


As the volumes of information available from modern sensors and sources increases, processing all available intelligence data has become too complex and time consuming for intelligence analysts. There is a need to assess how technology can and should be used to assist them, and for a clear view of the issues raised by the data and processing methods available. This paper discusses automated intelligence analysis, its impact and its role in modern conflict analysis.

1. Intelligence Analysis

Before we can automate intelligence analysis, we need to know what it does. The output of intelligence analysis (by human or machine) is intelligence: information that is pertinent to the current area of interest of a commander or policy maker, in a form that he can use to improve his position relative to another, usually opposing, commander. The other output of intelligence work is counterintelligence, misleading information supplied to the opposing commander (well-managed counterintelligence should also improve the commander’s position relative to his opponent). Simplistically, intelligence analysis is the collection and painstaking sifting of information to gain insight into and build a model of part of the world (a battlefield, country or other topic of interest ) that is relevant to a commander’s needs and actions; intelligence is a set of beliefs about the state of that world and the actions and intents of another party within it.

1.1 The need for intelligence analysis

To make decisions commanders need accurate and timely information based on as much relevant data as possible. The importance of good intelligence cannot be overstated; in a military situation, a commander is effectively blind without the battlefield awareness gained from intelligence. The view that a commander has of a situation depends on information pools and flows, as shown in table 1 and figure 1, where each OODA loop consists of Observe (using sensors and knowledge sources ), Orient (intelligence analysis ), Decide (command decision making using both learnt doctrine and individual style) and Act (move, oppose or contain a move). Note that in modern conflict, there may be 0, 1, or several other forces’ OODA loops in a commanders viewpoint, and that too little may be known about their structures, doctrine or equipment for conventional (orbat and doctrine-based) analysis. Given that each commander in a battle has their own view of it, it becomes important for a friendly commander’s battlefield awareness to be dominant (better than that of the opposition). Other considerations for intelligence analysts are that it also isn’t enough for analysts to provide a commander with information if it deluges him (the commander is given too much information to understand in the time he has available), changes too rapidly or he can ‘t trust it.

1.2 Why automate intelligence analysis?

Modern conflict focuses on the observation and control of other forces, where command and analysis relies on timely and accurate intelligence, Automated decision -making (currently at the research stage) will need fast, accurate sources of information to succeed. There is little argument about the need for intelligence in modern conflict, but before we consider automating intelligence processing, we need to decide whether, given the small scale uncertainty and high complexity of modern low intensity conflicts (Keegan’s guerilla-style real war), and operations other than war, it is a worthwhile thing to do. There are few metrics available for intelligence analysis: for human analysts there are the principles of intelligence, and intelligence processing systems are generally judged on the speed of their message processing and accuracy of their outputs (although accuracy is usually judged using a model of the world’s ground truth). These can be summarised by saying that intelligence processing should provide the commander with the most honest view possible of his area of interest, given the data, resources and time available.

Table 1 One commander’s information
* each force: World view including intelligence weaknesses, capabilities, intent, doctrine, deception, counterintelligence, available supplies and equipment, psychology, morale, coordination, allegiance,
* environment: terrain, possible changes (for example weather), human features (for example bridges and cities)
* situation: current state of play (situation and events), possible moves and events

Intelligence processing is a natural process for humans, but they have limits: when the input data size becomes too large to process within the time constraints given, or too complex (uncertain) to think clearly about, then automation of some of the processing must be considered. Given the increasing volume and complexity of information available from modern sensors, lower-level processing and open sources, it will soon be impossible (especially in a tactical situation) for analysts to process all of it efficiently and on time, and time now to start thinking about how automation can help. The second argument for automating intelligence processing concerns the advantages (mainly processing speed and uncertainty handling) that a reasonable intelligence processing system would give a commander. Concerns about the conflict superiority that technology gives are not new (indeed, they date at least from when the first arrow was made), but the emphasis has changed; recent warfare has moved from the production of larger, more efficient weapons to the observation and control of other forces (although this idea dates to Sun Tsu’s predecessors). This is reflected in the focus of current technology / research, at the Orient (intelligence processing) and Decide (decision making) stages in figure 1 (the progress of defence automation, as measured on the OODA loop). It is perhaps preferable to understand or possess intelligence processing systems than to fight a force that is better prepared by using them.

2. Models for automating intelligence analysis

The cognitive theory of intelligence analysis has been well studied. Not surprisingly, the intelligence cycle has much in common with models of human perception and processing. These are useful starting points in defining what is needed in an intelligence processing system. We learn that intelligence models are incomplete (we can never model the entire world), disruptible (models will always be vulnerable to external influences and counterintelligence), limited (models will always be limited by sensor capabilities), uncertain (input information and processing are usually uncertain) and continuously changing (models must deal with stale data and changes over time).
Requirements for future intelligence processing systems include rigorously handling uncertainty, partial knowledge (of inputs, area, opponent behaviour, training and equipment), detecting and cleanly retracting misleading information (including counterintelligence), giving credible and understandable explanation of reasoning; handling data that changes over time and including quantitative sensor -derived data in a symbolic reasoning framework. How much the computer should do, and how much should be left for the analyst is also an interesting problem. The analysis of and attempt to automate ‘judgement’ is a necessary step for the automation of low-level command decisions, but the change from assisting analysts to automated intelligence processing will redefine the operating procedures of military staff, and is expected to meet opposition. Since intelligence must be timely to be useful, the control of reasoning in these models must also be addressed : for example, since the creation of intelligence models is cyclic (as more intelligence is generated, so the gaps in the commander’s knowledge are pointed out, and more intelligence gathering is needed), when to stop processing and where to focus processing attention may be issues.

Decisions to be made include whether to distribute processing, and the acceptability of anytime algorithms (these output timely but less accurate information). To be useful, information must also be relevant to its user. Relevance is difficult to define; attempts to model it include user modelling and user profiling, and these should be allowed for in models if possible. Both the intelligence cycle and high-level data fusion models have been considered as a basis for intelligence processing models. These are not the only models that may be appropriate ; the design of generic intelligence models and processing should, as far as possible, incorporate techniques from contributing areas like cognitive psychology (mental models), detective work (psychological profiling), data mining (background information processing and user profiles) and data fusion (situation assessment ).

2.1 High-level data fusion

Data fusion is the combination of separate views of the world (from different information sources, times or sensors bandwidths) into a big picture of the world. Data fusion can occur at different levels of processing, from combining sensor outputs (low level) to combining processed information (high level). Intelligence processing, in its combination of information from different sources, is equivalent to and can gain from high-level (and also low-level) data fusion techniques.

2.2 The intelligence cycle

The Intelligence Cycle describes intelligence processing in stages. Three models are considered here : the UK and US models (which have different names for essentially the same stages), and Trafton’s model which adds an extra stage (utilization) to reflect use of the intelligence. The UK model with direction and utilisation stages added (as shown in figure 2), will be used in the rest of this section. The intelligence cycle isn’t just a description of the processes of intelligence, it also shows the flow of information within intelligence analysis, from commander to intelligence analysts and back. Within the cycle there are subcycles, the most important of which is the redirection of sensors, data collection to cover areas of ignorance found during processing, The rest of this section describes automation of each part of the intelligence cycle.

2.3 Planning, Direction and Collection, deciding what intelligence to collect

Information that analysts use to produce intelligence includes sensor data and intelligence reports, but there are other less obvious inputs : the information that a commander has requested ,both now and in the past, stored knowledge about the area of interest, and knowledge of opponents resources, training and expected behaviour. Input information is usually uncertain and often untrustworthy. Intelligence sources are classified into human, image, signals and measurement intelligence, the inputs from which are often labelled with uncertainties (source and information credibility for human intelligence, sensor errors for other categories of data). Since there is often a larger requirement for intelligence data than collection agencies can meet, there is scope for automating collection management with a constraint satisfaction algorithm, reasoning system using the priority and cost of data as inputs (Currently available tools include the JCMT system).

2.4 Collation and Analysis, processing the input information

Intelligence processing covers the stages of the intelligence cycle where intelligence is created from information. Processing by analysts is broken down into collation (sorting input data into groups, for example by units or subject),and analysis (making sense of the data and producing the big picture from it). Analysis is sometimes broken down into evaluation, analysis, integration and interpretation . Overall, intelligence processing uses input information to update knowledge about the situation, and create relevant intelligence reports. Intelligence processing doesn’t have to be very sophisticated to make a difference to an overworked analyst ; typical intelligence processing requirements are classification and matching reports to units, both of which are within the capabilities of current systems (although there is always some expected error). A good intelligence processing system should be capable of at least these two functions, with room for extension to techniques like recognising counterintelligence, behaviour and intent as research progresses. Intelligence processing is a large part of the reasoning in an intelligence analysis system, and has been given a separate (later) section in this paper.

2.5 Dissemination and utilisation, getting intelligence to the user

Intelligence is no use if it doesn’t get to the user on time and in a comprehensible, credible form : dissemination is essentially an attempt to adjust the user’s mental models of the situation – something which will not happen if the user does not trust the system. Utilisation is a catch-all term for using the data ; any actions taken by the user will affect the state of the real world and may change the commander,s intelligence needs, restarting the intelligence cycle (at the collection stage).
Automating intelligence processing will change the style and types of output available from it. Rigorously handling uncertainty in inputs and processing will improve the accuracy of intelligence processing outputs, but will also give the system designer a difficult choice between providing definite but inaccurate information to the user (this approach is preferred in, for example [10]) and more accurate but possibly confusing information (for example “T123 is a tank with 90 percent accuracy), or a set of possible explanations for the current data.

3 Automating Intelligence Proccesing

Table 3 shows some of the problems and issues in processing intelligence data.
The most pressing of these (uncertainty handling) is discussed further in [4].

3.1 Processing frameworks

Frameworks in existing intelligence processing systems include assumption-based maintenance systems, blackboards and graphical belief networks. Of these, belief networks seem most promising as a generic intelligence processing framework, since they can be used to combine uncertain sensor-derived data and knowledge within a probabilistic framework, and have a body of theory behind them that includes sensitivity analysis. The use of belief networks in intelligence processing is discussed in a separate paper; recommendations from it include further research into processing frameworks that includes study of :
* Focus of processing attention
* Handling of time-varying information
* Retraction of information and its ,traces, within the system
* Explanation of reasoning, including extraction of alternative explanations
* Multiple viewpoints and values
* Recognising and analysing group behaviour
* User profiling

Table 3 Problems for intelligence processing
* Input: multiple reports about the same event,object, information incest ,repeated and single,source reports, untrustworthy information, its removal and effects, varying credibility of information
* Representation: mixture of text,based reports and sensor data, several possible explanations for data, handling large numbers of parameters
* Speed: increasing amounts of complex information,data need for accurate, timely information combinatorial explosion
* Interface: credible and understandable explanation of reasoning, human inspection of data and intervention in reasoning
* Time: data that becomes irrelevant , deciding and removing it data and situations that change over time.

More than one framework could be used to split processing into essential and background work. This would provide opportunities for analysis of the patterns and flow of intelligence data, including dominant routes, behaviour and counterintelligence.

3.2 Existing intelligence processing systems

System designs and studies of how to automate Intelligence Processing are increasing in response to the need to take some pressure off the users and analysts. Although systems that handle intelligence data exist, some (for example GIFT, AUSTACCS) assist users to access information and don’t attempt processing; many (for example ASAS, DELOS) do not process it beyond the correlation (sorting data into groups) stage of the intelligence cycle, and most are vulnerable to uncertainty in their inputs and missing data in their knowledge of tactics and equipment. Systems that attempt full intelligence processing (for example HiBurst, IMSA, TALON and Quarterback) are usually based on Artficial Intelligence methods which include assumption-based truth maintenance systems, argumentation and expert systems augmented with fuzzy logic. Although probabilistic networks seem a promising base representation for intelligence processing,
research on intelligence fusion using probabilistic networks appears to be confined to teams at the UK Defence Research Agency, George Mason University and some US commercial sites.

3.3 Using Data Mining as a testbed for Intelligence Processing

There is a vast amount of electronic data even in a single company. Most of it is distributed across several systems /formats, and is useful but inaccessible. Data Mining is the process of extracting implicit and relevant information from data sources, which are usually databases, but can be open sources ,e,g, the internet. Data Mining gives us the opportunity to test intelligence fusion algorithms using open source information as input.

4 Automating counterintelligence

As counter-terrorism and infiltration units have discovered, getting inside the opponent’s OODA loop is more subtle than destroying communication links : his information can be manipulated to our commander’s advantage. We can view his Observe process using stealth and EW technology : we can disrupt it using deception. Counterintelligence also affects the Orient and Decide processes, both by data deluge of processing resources and the creation of uncertainty in an opposing commander’s mind. The creation of models of an enemy commander from his known doctrine and reactions also allow us to anticipate his moves rather than just react to them – it also makes the targeting of counterintelligence (for instance, in information warfare) possible. This augments the current countermoves of defending against enemy actions (e.g. using air and ballistic missile defences). As conflict can be viewed as an interacting series of these pairs of OODA loops, this is also a useful starting point for the automation of low-level command decisions.

5 Conclusions

Intelligence processing creates a belief in the state of a world (ie a battlefield) from uncertain and often untrustworthy information together with inputs received from sensors. This is a natural process for humans, but they have limits : when the input data size becomes too large to process within the time constraints given, or too complex (uncertain) to think clearly about, then automation of some of the processing must be considered. This paper outlines issues to be considered in designing the next generation of UK intelligence analysis systems. Future systems should allow analysts to concentrate on high-level analysis rather than clerical operations like duplication elimination. Automated intelligence techniques and systems are being designed in response to this need. Most of them assist analysts by providing better information retrieval and handling tools. The automated processing of intelligence is beginning to be addressed, but lacks a mathematically sound representation ; techniques based on data mining, cognitive psychology and graphical networks are promising, but need further research effort. This work also unites sensor-based data fusion systems with knowledge-based intelligence processing and decision support. Possibilities arising from using a complete and efficient representation include the ability to use most of the information available to a system, the analysis of patterns of behaviour and the generation/ recognition of counterintelligence data.


1. Nato Intelligence NATO report AINTP
2. Canamero D Modeling plan recognition for decision support European Knowledge Acquisition Workshop
3. Companion M A and Corso G M and Kass S J Situational awareness an analysis and preliminary model of the cognitive process Technical report IST TR University of Central Florida
4. Farmer S J Uncertainty handling for military intelligence systems WUPES Prague
5. Katter R V and Montgomery C A and Thompson J R Cognitive processes in intelligence analysis a descriptive model and review of the literature Technical report US Army Research Institute for the Behavioural and Social Sciences
6. Keegan J A History of Warfare Random House London
7. Keegan J Computers can t replace judgement Forbes ASAP December
8. Laskey K B and Mahoney S and Stibio B Probabilistic reasoning for assessment of enemy intentions Technical report C I, George Mason University
9. Shulsky A N Silent Warfare understanding the world of intelligence Brassey s US
10. Taylor P C J and Strawbridge R F Data fusion of battlefield information sources Royal Military College of Science Shrivenham
11. Trafton D E Intelligence failure and its prevention Naval War College Newport RI
12. Sun Tsu date unknown The Art of War Oxford University Press

Old papers: Uncertainty Handling for Military Intelligence Systems

Something I had laying around the office; thought it would be useful for the archives in a “what i thought when I was 10 years younger” sort of way.


We describe sources of and techniques for handling uncertainty in military intelligence models. We discuss issues for extending and using these models to generate counterintelligence- recognise groups of uncertainly-labelled entities and recognise variations in behaviour patterns.


Intelligence is the information that a commander uses to make his decisions. It is an informed view of the current area of interest of a commander or policymaker, in a form that he can use to improve his position relative to another, usually opposing, commander.
Uses of intelligence include the basis for command decision making and the creation of uncertainty in opposing commanders’ systems and minds. Commanders use intelligence to recognise situations (situation awareness), predict changes in situations, predict an enemy’s behaviour (threat assessment) and decide which actions to take (planning) The quality and availability of intelligence (rather than information) determines whether a force is reactive (can only react to its environment or opponent’s moves) or proactive (can make informed plans and manipulate its situation).
Two major problems for commanders in the Persian Gulf conflict were the volume and complexity of intelligence data. If these are to be alleviated, methods for producing efficient representations of input data and information must be found – this includes automating the processing of raw intelligence data into useful knowledge


Know the enemy and know yourself in a hundred battles you will never be in peril, When you are ignorant of the enemy but know yourself- your chances of winning or losing are equal, If ignorant of both your enemy and yourself- you are certain in every battle to be in peril’ (Sun Tsu)

The role of intelligence processing is to make sense of the world by piecing together the uncertain, conflicting but usually copious evidence available
Intelligence is information that is pertinent to the current area of interest of a commander or policy maker, in a form that he can use to improve his position relative to another, usually opposing, commander. Although some intelligence work is the stuff of James Bond and Le Carre novels, intelligence analysis is painstaking sifting of information to gain insight into the actions and intents of another party Intelligence processing creates a model, or informed belief, of the state of that part of the world which is relevant to a commander’s decisions and actions. Intelligence is produced by fusing uncertain and often untrustworthy information (sensor outputs and text-based reports) with prior knowledge (e.g. enemy equipment and tactics). This is a natural process for humans, but when the input data size becomes too large to process within the time constraints given, or too complex to think clearly about (people do not reason rationally under uncertainty), then automation of some of the processing must be considered The flow of analysis is usually based on the intelligence cycle : Direction – deciding what intelligence is needed; Collection – collecting information; Collation – sorting information; Evaluation – processing information into intelligence; and Dissemination – giving that intelligence to the commanders/ users (NB this is the UK definition of the intelligence cycle; different labels are used in the US definition)
Current intelligence is gathered on a limited number of topics or geographical areas, but currently irrelevant basic intelligence is also processed and stored ready for when the attention of a commander or situation shifts. The creation of intelligence models is cyclic; as a better picture of the world is generated, the gaps in the commander’s knowledge are pointed out, and more intelligence gathering is needed (this is shown in the diagram below by a feedback loop from collection to evaluation).
Characteristics of intelligence systems are that they are driven by a set of goals (the commanders requests for information), have information sources that can be partially controlled, situations that change over time and a large body of input information that is uncertain and incomplete. There is normally at least one non cooperating red agent capable of actions against the blue commander using these systems. Enemy actions against blue’s intelligence operations include counterintelligence and deception: attempts to distort blue’s model of the world Other agents that may need to be modelled include neutral forces and civilians Intelligence processing concentrates on resolving the uncertainties caused by inaccurate, infrequent and incomplete inputs, cultural differences, counterintelligence and approximate reasoning.
Intelligence models are :
* incomplete (we can never model the entire world),
* disruptible (models will always be vulnerable to external influences and counterintelligence),
* limited (models will always be limited by sensor capabilities),
* uncertain (input information and processing are usually uncertain) and
* continuously changing (models must deal with stale data and changes over time)
Military intelligence can go one step further than just modelling an uncertain
World; in using counterintelligence and deception about his plans, situation and
actions, a commander is creating uncertainty in an opposing commander’s models
The use of counterintelligence is one of the main differences between military intelligence and other uncertainty handling models (although there are similarities in handling counterintelligence, fraud, input errors and cultural differences).


Although intelligence is currently processed by analysts, its automation is being driven by increasingly smaller time-frames, greater volume and complexity of available information Intelligence processing is increasingly similar to high-level data fusion (intelligence level fusion); making sense of the world from as much input data and information as possible.
Military conflict is essentially chaotic It is a sequence of well-defined moves that interact locally, yet produce long-range effects At the local level it is still possible to model these effects if they are bounded by physical laws, resources and the trained behaviour or rules of the parties involved
During the cold war, the West faced known enemies on known territory with well modelled outcomes (a winter war across Germany) Post cold-war intelligence analysis deals with more uncertain (less is known about the enemy) and complex (conflict is more likely to be in a setting which contains neutral populations) environments and forces Although small-scale, terrorist and guerilla conflict may seem random, they are still constrained (by environment and logistics), their players still trained (often in tactics well known to the west) and their sequences of actions still partially predictable.
Automated intelligence analysis systems are limited by time constraints and are unlikely to produce perfect summaries of the world It should be stressed that their prime function should be to improve current intelligence analysis The aim of this work is not to produce exact solutions and assessments of uncertain inputs and situations, but to give a commander as honest an assessment of a battlefield as possible within the constraints of the inputs, uncertainties and processing time available This paper focuses on the sources of and methods for handling uncertainty in military intelligence systems: [5] discusses other aspects of automating intelligence processing in greater depth.
War is the realm of uncertainty three quarters of the factors on which action in war is based are wrapped in a fog of greater or lesser uncertainty, the commander must work in a medium which his eyes cannot see which his best deductive powers cannot always fathom and with which- because of constant changes- he can rarely become familiar [4]
Uncertainty is not an important issue most of the time, as a commander will recognise the situation and react to it. Issues to be addressed include sources of uncertainty, whether we can improve our sensor allocations to reduce uncertainty, and how much uncertainty matters (how much uncertainty we can tolerate before a system is ignored or useless).
An intelligence processing system should use all (or as much as possible) of the information available to it This information is more than just input reports and sensor data; the context of an operation, open source, analysts’ knowledge and the needs, preferences of users are also available. The systems should not take every input fact as certain; fortunately, most of this information is tagged with source and information credibility, sensor accuracy or a range of possible values.
The reasoning framework used cannot be divorced from decisions about how to handle uncertainty The aim of an intelligence processing system is to use prior experience and knowledge to pull out the information implicit in input data, whilst losing as little of that information as possible One of the main differences between reasoning frameworks is the point at which they discard information This ranges from rule-based expert systems, which force a user to decide on the truth or falsity of input statements, to systems which manage uncertainty about inputs, conclusions and reasoning to produce an assessment of a situation which takes account of all of these This latter system is most desirable.


Counterintelligence is the main difference between uncertainty handling in military and other systems Modelling a military domain is compounded by an enemy attempting to deceive sensors and subtly change our models of the situation Counterintelligence manifests itself as conflicts between conclusions and unexpected lack of accumulation of supporting evidence Conflicts can be traced back to sources and information and counterintelligence hypotheses included and evaluated This can be incorporated from the outset by regarding inputs as observations of hidden information (either intelligence or counterintelligence).


The information output includes physical data (geography and positions; movements of forces), tactics and expected behaviour patterns, and social factors. Although most of these should have uncertainties associated with them, they currently do not, and one of the first questions in building an intelligence processing system should be whether this matters and if so, how much The final point at which information is discarded (uncertainty occurs) is in the user’s mind Knowing what the user is interested in (user profiling) can focus the output Even if an honest summary of the situation has been produced, complete with uncertainties; probabilities of different scenarios and actions, if this model is not transferred to the user’s model of the world, then the processing will have been useless Users also suffer from hypothesis lock in which alternative explanations are rejected regardless of accumulating evidence Managing this phenomena requires good explanation of reasoning, uncertainty and evidence.


The choice of reasoning framework is central to this work, both in its flexibility and its handling of uncertainty. Although intelligence is currently processed by human analysts, attempts to model it have included fuzzy logic, belief networks, assumption-based truth maintenance systems and rule-bases with exceptions. A Belief Network is a network (nodes connected by links) of variables that probabilistically represents a model – ie the beliefs that a user has about a world. Its main use is as a reasoning framework for manipulating and combining uncertain knowledge about both symbolic and numeric information. Belief networks can be extended to make decisions based on a users’ stated preferences. Such networks are known as Influence Diagrams There is a large body of research into many aspects of their use which includes learning networks from data, temporal (dynamic) networks and efficient evidence propagation We consider Belief Networks to be an appropriate framework because they handle uncertainty in a mathematically rigorous way, and they can be manipulated to provide more than just a model of a world Our experience in using belief networks for such a complex and uncertain application has, however, highlighted shortcoming in current belief network theory Key problems identified include the lack of high-level structure, treatment of time-varying information (including hysteresis effects), correlation between the real world and the model, slow speed (we may need to accept tradeoffs between uncertainty and execution times), handling of ignorance, and their single model (viewpoint) of the world.


Analysis of typical intelligence problems has shown information to be hierarchical (and sometimes fractal), grouped and layered. An example is air picture compilation where an aircraft can carry several different weapons (which are each applicable to different types of target), and aircraft of different types are grouped into packages which then perform single missions. We propose the use of an object-oriented framework, where each object (i e aircraft) contains a network that can inherit nodes, sub-nets and conditional probability tables from a class hierarchy Each object network contains hooks – nodes that correspond to similar nodes in other objects’ networks Links between these nodes are often simple one-to-one conditional probability tables, but can be more complex; for instance, a package will have a one-to-many relationship with several aircraft This allows the dynamic creation of large networks from components It also allows the use of default propagations across objects (which are often single nodes in higher-level networks), default sub-networks (prototypes) and extra functionality (for instance the modification of input data). Using these robust architectures should improve network design times, but some consideration needs to be made of how much representation accuracy is lost in using them (for example whether adding an extra child link to a node will change the importance of its other children proportionally) Much of the theory has already been covered in discussions of semantic networks, plates, meta-nodes and representing conditional probabilities as networks We propose the use of self-organisation to create the boundaries between sub-nets, and the use of constraint satisfaction techniques to decide which hooks should be joined.


Intelligence processing is real-time and computationally expensive Ideas for overcoming the time-constraints and bottlenecks caused when processing large amounts of data include distributed processing, using hierarchical architectures to limit the spread of information, and modifying analog radial basis function chip designs to belief network representations We propose limiting propagation by collection of information at the boundaries of meta-nodes, approximate propagation across these boundaries, then propagation of batches of information properly when time allows.
Propagation can thus occur at the node or metanode level When propagation is allowed to proceed at both levels simultaneously (this is equivalent to using two layers of networks- one deterministic/approximate, the other detailed/probabilistic) the output will reflect the most detailed model for the time and attention constraints.


How a network corresponds to the real world, particularly the pragmatic and semantic subtleties of representing evidence uncertainty and ignorance, is also interesting. The problem of unreliable witnesses is so rife in intelligence processing that all information and human sources have reliability estimates attached to them. Current attempts to model this partial ignorance include using techniques from possibility theory to handle vague inputs, and using evidence nodes to spread the probabilities at input nodes.


Other issues that have been identified which impact on the automation of intelligence processing are:
* representation of time-varying information feedback (for instance using recurrent belief networks and methods based on Markov chains)
* incremental build-up of errors from evidence removal (rebuilding networks using only currently available data),
* multiple space and timescales (no current solutions, but some signal processing theory may help),
* multiple utilities (multiple attribute utility theory)
* when to refer problems to human operators (sensitivity analysis to data and data flows)
* multiple viewpoints to give a spread of possible outcomes rather than a point view of the environment (layered networks to avoid repeating entire networks
– see the section on real-time processing)
* reasoning about limited resources (colored network theory)
* discovering high-level patterns and trends in information, including behaviour patterns (adapting numeric pattern processing techniques to use symbolic inputs)
* generating behaviour novel plans (destabilising the networks – cf chaotic net theory)


Since any view of an environment is subjective, limited by the knowledge and information available, that view is open to manipulation by an intelligent adversary. This is the basic premise of information warfare; the planning of counterintelligence and deception moves (i e mock-up tank emplacements) to manipulate or attack a red commander’s mental model of the situation Information warfare is a powerful technique which complements existing command and control warfare (the disruption of communications between the red commander, his forces and intelligence). We already have models of blue’s view of a situation Some theory already exists for the adjustment of network-based models to their input/outputs, and for multiple views of the same situation It is therefore useful to adjust a blue model of a situation to create a blue estimate of the red commander’s viewpoint, using red’s known doctrine, sensors and reactions Sensitivity analysis of blue’s red commander model can then be used to determine which of several possible deception moves by blue would be most likely to alter the red commander’s view of a situation to that desired by blue.


The analysis of conflict, like game theory, embraces any interaction between parties with differing and usually contradictory aims Intelligence analysis provides a viewpoint from which an agent or human can decide and act in the real world. Applications of intelligence processing techniques range from battlefield awareness to security systems and intelligent data mining; our example/test applications include classifying combat aircraft missions from sensor data and recognising criminal behaviour patterns.


Intelligence processing is an interesting area for the application of uncertain reasoning techniques The main difference between this and other applications is the deliberate creation of uncertainty (counterintelligence) both by own and opposing agents. This gives a new perspective on uncertainty – that of a useful thing to create.


1. NATO Intelligence Doctrine NATO report AINTP-1 1996
2. AN Shulsky Silent Warfare Brassey’s US 1993
3. Sun Tsu The Art of War Oxford University Press
4. C von Clausewitz On War Princeton University Press
5. S,J Farmer Making Informed Decisions Intelligence Analysis for New Forms of Conflict IMA Conference on Modelling International Conflict- Oxford ),) April
6. W Feller An Introduction to Probability Theory and its Applications Wiley-
7. DA Norman and DG Bobrow On the data,limited and resources, limited processes Cognitive Psychology
8. R Szafranski A Theory of Information Warfare Preparing for, Airpower Journal- Spring
9. A Tversky and D Kahneman Judgement under uncertainty heuristics and biases, SIAM Journal on Computing
10. G Shafer Savage Revisited SIAM Journal on Computing
11. C Elkan The Paradoxical Success of Fuzzy Logic IEEE Expert- August
12. SG Hutchins and JG Morrison and RT Kelly Principles for Aiding Complex Military Decision Making Command and Control Research and Technology Symposium-Naval Postgraduate School- Monterey- California- June
13. J Pearl Probabilistic Reasoning in Intelligent Systems Morgan Kaufmann
14. RE Neapolitan Probabilistic Reasoning in Expert Systems Wiley
15. E Horvitz and F Jensen Uncertainty in Artificial Intelligence Morgan Kaufmann

What’s the questions again?

Oh the questions, the questions. The more meetings I go to, the more I see the deep fundamental importance of asking and attempting to answer the right questions. And there are many wrong questions out there, especially (but definitely not limited to just) in AI. Having said there are wrong questions, I probably need to clarify that. There are very few invalid questions: most questions provoke a response, or thought, and therefore have a motive or reason. Now some questions are just plain immoral (such as “what’s the best way to kill a million people”), others are just plain insensible (“Hey you! The troll with the big bike! How come you’re so ugly?”) and some are just there for the fun of it. So also having said that there are wrong questions, I need to qualify again that it’s a subjective thing. There are good questions to ask here and now, and there are bad questions to ask here and now. I’m just hoping that I have enough good taste to tell the difference.

So: questions to think about sometime on this blog, with the proviso that some of these questions are just for fun, with no promise of anything even approaching a bounded sensible answer:

* How can we organise machine thought?
* How is machine thought similar to and different from human thought?
* How can we make machines more creative?
* How do humans make sense of information?
* What do we do when we’re not sure?
* What makes us human?
* How are some people geniuses, and can we replicate that with machines?

Sometime, I may try to arrange these into the possible and not here/not now categories. But that may have to wait until after I’ve attempted to start answering them.

The city needs to go green?

No, not green as in environmental (although that would be a good idea too); green as in look to developments in defence for where it might go next.

Parts of the city seem to be buying Bayesian statisticians as if they were going out of fashion (they’re not: from what I’m hearing on the tech grapevines, belief is apparently the new black for this season). Which makes sense in a world where predictions are shifting from ‘did we see something like this before and what happened’ to ‘we haven’t got a clue what’s happening next and the best we can do is guess from fragments’. I spent a day moving on a bit from that, playing with situation awareness techniques, and wondered if that was where the city might go next. Only time will tell…

Structure Mapping Theory

More old notes to be edited later… so many of these to be done…

In analogy, the relations between objects (e.g. friend(Bill,Ted)) are matched rather than the individual attributes (e.g. tall(Bill)) of those objects. This is distinct from literal similarity where both attributes and relations are matched, mere-appearance matches where primarily attributes are matched, and anomaly where very few attribute or relations are matched. Metaphor (e.g. ‘Bill is a rock’) is seen as a reduced form of analogy with usually just one (contextual?) attribute being matched. Abstraction is seen as a similar process to analogy but with mappings between objects that have few or no attributes, and in abstraction, all the relations are matched rather than just some. This view of analogy appears to be confirmed by empirical psychological studies (Falkenhainer89).

Analogy can be split into three distinct subprocesses (Falkenhainer87.IJCAI): (i) access, (ii) mapping and inference and (iii) evaluation and use. Before processing begins, it is assumed that there is a current situation of interest (the target). Access finds a body of knowledge (the base) that is analogous or similar to the current target. Mapping finds similarities or correspondences (mappings) between the base and target, and may also transfer information (inference) from the base to the target. Evaluation of the analogy gives a measure or estimate of the quality of match in the context of other knowledge about the target’s domain, against three different types (structural, validity and relevance) of quality criteria. Structural criteria include the degree of structural similarity, the number of similarities and differences and the amount and type of knowledge added to the target description (the candidate inferences) in using this analogy. Validity checks that any new knowledge makes sense within the current knowledge domain. Relevance assesses whether the analogy and candidate inferences are useful to the current task or aim of the system.
Analogies thus formed are used for analogical reasoning, similarity-based generalisation, or analogical learning.

Structure mapping theory (SMT, Gentner83) asserts that an analogy is the application of a relational structure that normally applies in one knowledge domain (the base domain) to another, different, knowledge domain (the target domain); unlike less-structural psychological theories, it also sees analogy and similarity as connected processes. It uses graph-matching to describe the constraints that people use in creating analogies and interpreting similarity. SMT does not capture the full richness of analogy, but is primarily interested in, and applied to, the mapping and a subset of the evaluation subprocesses described above (e.g. it does not include the access subprocess, and uses structural criteria only).

The core of structure mapping algorithms is finding the biggest common isomorphic pair of subgraphs in two semantic structures [Veale98]. Much of the work in this field has concentrated on subgraph heuristics to produce near-optimal matches in polynomial time. All of it is based on or related to SMT. In SMT, graph matches are assumed to be exact correspondences only (structurally consistent); this constraint may need updating to better reflect the inexact matches and representations used in human analogical reasoning. Note that there are two matching processes in play here: the mapping of objects (graph nodes), and the mapping of relations (link labels), that these two processes are often run at the same time, and that the relational matches may determine the matches between objects. SMT is also tidy: attributes are discarded unless they are actively involved in a match, and higher-order relations and interconnections are more likely to be preserved (systematicity, Gentner83).

SMT underlies much work on computational models of analogy, metaphor, case-based reasoning and example-based machine translation [Veale98]. SMT is used for analogical learning (Falkenhainer87.IJCAI, JonesLangley95) and information retrieval. Analogical learning uses analogy to generate new knowledge about a domain; examples include the extension of the ARCS/ACME structure-matching programs with algorithms from the Eureka learning system. Similarity-based retrieval models include MAC/FAC (many are called but few are chosen, Gentner95); this is a two-stage algorithm, where the MAC stage is a rough filter, and the FAC stage is performed by a structure-matching engine (SME).

Implementations of SMT include the Structure Mapping Engine (SME, Falkenhainer89), SAPPER, ACME and ARCS. SME represents each objects as a description group (dgroup), a list of items (entities and expressions) associated with that object, where functions, attributes and relations within these items are represented as Prolog-style clauses. Each dgroup is mapped onto a graph with a node for every item in it, and an arc from every item i to any item j which is used as an argument to i.
The SME algorithm is divided into four stages: Local match construction, Gmap construction, Candidate inference construction and Match evaluation. A local match is a pair of base and target items which could potentially match; in local match construction, a match hypothesis is created for each of these pairs. SME does not create a match hypothesis for every possible pair of base and target items, but uses a set of rules (for instance, only matching items that appear in the same expression) to filter them. Which rules are used here determine whether the output of SME will be analogy, literal similarity or mere appearance. Because they are based on dgroups, match hypotheses also form a directed acyclic graph, with exact matches fixed only between predicates. GMAP construction is the heart of SME processing. Here, the local matches are combined into gmaps, where a global mapping (gmap) is a structurally consistent mapping between the base and target object, consisting of correspondences between the items and entities in their description groups, candidate inferences and a structural evaluation score. Candidate inferences are the structurally grounded inferences in the target domain suggested by each gmap, where the substitutions used in structurally grounded inferences are consistent with the existing target information, and there is an intersection between their ancestor nodes and the network representing the target domain. In match evaluation, a score is given for each local match, using a set of rules (for instance, expressions with the same functors are assigned a score of 0.5). These scores are combined using Dempster-Shafer theory (in a Belief Maintenance System) to give a score for each match hypothesis. The scores for all the match hypotheses that contribute to an individual gmap are then summed to give a structural evaluation score for that gmap.

Veale’s SAPPER system [Veale98] represents each domain as a graph in which nodes represent concepts (e.g. SME’s items) and edges between nodes represent relations between those concepts. The algorithm uses much simpler rules than SME. It first looks for common nodes between the two graphs being compared, then uses two simple rules to update the match: the triangulation rule and the squaring rule. This creates a partial match pmap for the domain, and is a spreading-activation algorithm that Veale claims is more efficient than the corresponding part of SME [Veale97]. The pmaps formed are graded by richness, then the algorithm attempts to combine all the pmaps, starting with the richest as a base.

Contenders to SMT for modelling analogy include high-level perception (Hofstadter’s CopyCat) and theories of semantic/pragmatic constraints. High-Level Perception emphasises analogy as a mixture of representation and mapping, and treats data gaps in a similar manner to recent work on protein structure alignment. Holyoak and Thagard’s work on the Analogical Constraint Mapping Engine (ACME) and ARCS systems assume that analogy is goal-driven; Holyoak [Holyoak89] argues that the SMT view of analogy only addresses the structural constraints on a mapping, and that semantic and pragmatic constraints should be included in the mapping too.

Where an analogy is drawn between two separate areas of knowledge, then the use of information fusion by them is simple: the joining of the information held in the two pieces of knowledge into a coherent whole. More complex forms of creativity combine together several pieces of knowledge; this is where information fusion techniques will be most useful and provide most insight into the process.


Is insight creativity applied to thought processes? I mean, if art is creativity using materials, and maths creativity using symbols, can we distill the notion of insight right down to creativity using partial world views? It could fit: the idea of generating then analyzing and reducing sets of ideas down to ones with higher value (for a given notion of value), of combining partial views, refining them and creating leaps by adapting or removing base assumptions. So are the people who spot the moving patterns and understand trends fast enough to profit from them creative (as they’re adapting their models quickly) rather than purely logical. Possibly. Could we use structure-mapping to model those processes? Some, but that would definitely need deep parsers, large knowledgebases and one heck of a legal framework…

When we created the 4-tier fusion pyramid (data-information-knowledge-insight/wisdom; you’ve been using it for years chaps, and now you can finally credit Mark Bedworth. Who isn’t me, btw…), I placed insight on the top tier. It made sense as a simple analogy; if all the raw materials in intelligence processing could be heaped together, then the distillation of a larger amount of knowledge into a smaller amount of insight could be viewed in a similar way to the distillation of a larger amount of data into a smaller amount of information. It also tacitly acknowledged that the representation of insight might be quite different to that of knowledge, that somehow (as with data to information), the linkages within and context of that knowledge could be used and shown more concisely. And now I need to think very very carefully about what exactly I meant by that. But first, I need to sleep. Goodnight.

Thinking Simply

I’m not a very intelligent person, but I am lucky enough to know several of these creatures. I got to wondering this weekend about what, apart from the ability to assimilate and use vast amounts of information faster and more efficiently than us mere mortals, really sets them apart.

And the thing that strikes me most is their ability to think very very simply. Now I’m not talking about thinking simplistically, i.e. with the sort of logics found mainly in the Daily Mail or the town pub after 10 pints. I’m talking about the ability to take a topic, a known topic, to reach into its core and retrieve an idea that seems so simple, so obvious, that you wish you’d thought of it yourself. And that little inner voice tells you that you could have done, if only you’d noticed, but you know in your heart that only someone very smart has the gift of thinking that simply.

And then I started wondering how this simplicity might be connected to my long-ago quest to understand creativity; if what the smart people were doing was having the courage, confidence and tools to rearrange existing concepts that build up around a topic over time. Which took me back to an old book (the structure of scientific revolutions), a coarse precis of which is that science moves at two speeds: slow methodical progress interspersed with great leaps of the imagination.

Meanwhile, I’ve been reading some of my old notes and, once again, can’t understand them. Context is all sometimes…


One of the things that occasionally fascinates me is how we define ourselves as human by defining other animals as somehow, well, inferior. And one of the areas that we’ve traditionally done that til recently is definig ourselves as the only toolmakers. Now anyone who’s ever looked at an empty snail shell with a hole in the side without going “wow, I never knew they had escape hatches” automatically knows that isn’t true. But how untrue? Once more to the literature boys…

Several books and articles (e.g. this 1940s article by Kenneth Oakley) on man the toolmaker take toolmaking as a given; these works move straight to which tools for what purpose, and skip the reasons and methods by which we might have become toolmakers at all. There’s a species of early man, homo habilis (“able man” or sometimes “man the toolmaker”) named after its toolmaking skills, mainly for an ability, 2.6 million years ago, to create and use a cutting edge on a small stone.

I’ll probably come back to people later, as toolmaking appears to be bound up with creativity, definition and even religion. For now, I’ll concentrate on the animals, as a useful control group. So, the ones that I know of are:

And lots more behaviours being observed in labs. There are even behavioural ecology groups out there studying this. In the end, this is really a non-post because the arguments are all there, the evidence is known, including that animals will use whatever materials are to hand (or claw or beak) for the tasks that they have, and that well-fed, well-cared-for captive animals have more time and inclination (even with adjustments for other factors) to be creative; the questions really are only of degree. And that’s before we talk about elephants and primates painting.

And then there is the use of animals as tools by humans. Which is a whole different subject, and not one for today. Suffice to say that the Mk 7 Dolphin does exist, and even has patents attached.

Defining Creativity

Sometimes I’m going to post some old notes here. This is one of them.

There are many different attempts to define creativity, and work on automating creativity is often inhibited by the au­thors’ own definitions. The process of creativity has been divided into several stages. Hadamard’s description of Poincare’s four stages is used by Boden amongst others: these are

  • preparation ­ define the problem, and attempt to solve it by rational means.
  • incubation ­ generate novel patterns or concepts.
  • inspiration ­ recognise a creative solution.
  • verification ­ rationally compare the solution to the problem.

The preparation and verification stages may not exist be­cause there may not always be a given aim to creative work. Incubation and inspiration are however central to creativity: it always contains a two­ part process of generation of concepts then evaluation of how creative those concepts are.

Defining the problem

The act of finding a problem is usually part of the creative process. Some creativity systems are very focussed on this: for example, in flexible means­end analysis (Jones+Langley) the problem is defined as a current world state and a set of goal conditions. This part of creativity is very closely related to conven­tional learning theory: Thornton (C.Thornton, 1998) has argued that the bias inherent in any recursive learning algo­rithm can be viewed as a form of creativity.
Although I have stated that preparation and verification are not necessarily essential to creativity, they are very impor­tant: perhaps the difference between creativity and random­ness, between human creativity and madness is in its con­nection to a purpose or communication (for example, even in describing art, we speak of its expression).

Generating novel concepts

Three main types of creativity are Boden’s (M.Boden, 1990) improbabilist and impossibilist creativity, and a chaotic form of creativity seen in many of the neural­ network based approaches.

  • improbabilist creativity is the construction of new concepts from existing ones, often combining previously­unconnected information to solve a previously­ unseen problem. The lightbulb puzzle (including the information that lightbulbs are hot just after they are on) is an example. Improba­bilist creativity was explored by Koestler (his ‘bisoci­ation of matrices’ (A.Koestler, 1964)) and discussed by Perkins (DN.Perkins, 1981).
  • impossibilist creativity transforms the space in which a concept can exist. This includes widening the frame of information around a concept being examined, and the removal of assumptions or constraints from the en­vironment in which a concept exists. Jackson Pollock putting his paintings on the floor (removing the as­sumption that paintings need to be vertical) is an ex­ample of this. There are many constraints at play in creation. For in­stance, in the creation of prose, the pattern of stresses in a line is as important as the meanings and rhymes and hidden meanings within a stanza. We work within unspoken rules: creativity can work within these rules (using them as guidelines) or on those rules them­ selves (to create new forms or categories of art or sci­ence).
  • chaotic creativity is where a small mutation of an ex­isting concept is allowed. Beethoven’s minute rework­ing of his musical themes until he hit one which was acceptable to him, and Thaler’s creativity machine are examples of this.

This seems a reasonable division to work with, although it would be interesting to see whether, when these three forms of creativity are finally modelled, other forms of creative act and process become apparent.

Measuring creativity

Creativity is often confused with the creation of new things. Creativity is not novelty: while generation of concepts is important, it is not effective without their evaluation. Evaluation consists of deciding which solutions are cre­ative, either by clustering them or by using a measure of surprise. To be creative, we need some sense of the differ­ence between what is truly creative and what is just new. We need to have a sense of how to cluster the mutations generated and how to define the boundaries between those clusters: we need a sense of taste or discrimination. Much of this, we can take from work on concept clustering and in­formation fusion, and work on the difference between cre­ative solutions and novel near­miss solutions to a problem, and the change in process that leads to them. As an example, take the humble paperclip. I can bend a pa­perclip into dozens of minutely different and new shapes, but only a few could be seen (without an explaining con­text, which is in itself a creative act) as creative mutations of its original shape.

Comparing the solution to the problem

If the creative process is used to solve a specific prob­lem, then the problem and its potential solutions need to be matched. Again, this process is closely related to the process of assessing the output from conventional learning algorithms.