CHAPTER 21

Modelling Motivation

21.1 .An Exercise in Robotics

The complexity of human emotions makes the thought of trying to model these in any explicit way a daunting task. The complexity may, however, be sorpewhat reduced by taking a rather unflattering and simplistic view of human nature. The result may be a form of representation which ignores many of the subtle nuances of human nature, but it will carry our system surprisingly far. In short, the human organism can be viewed as a gratification-seeking automaton.

Imagine that we have been asked to design a robot which will later be sent to some distant planet of which we have almost no knowledge. The environment in which it will find itself, and in which it must survive, may exhibit very strange phenomena-and could be populated by all kinds of strange and possibly hostile life-forms. Let us assume that technology is available so that we can provide the robot with external sensors, eyes, ears etc., with which it can perceive its surroundings. However, we have no way of knowing how to program it to respond to the things it sees, hears, touches etc. since we do not know what they are or how they will behave. A strategy which we might adopt to solve this problem is as follows.

We can provide the robot with internal sensors. These can detect its oil pressure, its temperature and whatever else is appropriate to the innards of a robot. With these it can detect its inner condition. At least we know what these conditions should be for survival, and we can program it so that it can recognise when it is in a desirable condition and when it is in an undesirable condition. We will call these 'goals' and 'anti-goals' respectively. We can also arrange its sensors so that it will be able to detect when its inner conditions are changing for the better or the worse (towards goal state or towards anti-goal).

Next we provide it with something like a 'black-box flight recorder'. This records all its external and internal perceptions as they occur, and stores them together with indicators to show when the sequence of events recorded led to goal and to anti-goal conditions. Of course, events are unlikely to repeat themselves exactly, but if the robot can process and analyse its recordings so they can be classified by similarities, ignoring circumstances peculiar to each unique event, then there is some chance that it will be able to recognise history repeating itself and take action to select favourable outcomes and avoid unfavourable ones. From these records the robot could form for itself sequences of hypothetical but probable events, in order to predict the likely outcome from known starting conditions.

The formation of these chains could become too complex and time ­consuming for real-time computation, and so it would be expedient to develop sub-goals which were known from previous experience to be conditions from which the main goal state was easily achieved. Some way would need to be found to label these sub-goals as 'desirable' in their own right. It would, for the same reason, be expedient to identify sub-anti-goals which should be avoided whenever possible. Note that we cannot identify the sub-goals in advance, since .here on Earth we have no way of knowing what they will be. What we must arrange for is an automatic labelling of the sub-goals after they have been discovered. .

Some behaviour patterns could be pre-programmed into the robot. These would be basic behaviour patterns such as that required to maintain its fuel supplies. For example, if the robot could recharge its batteries by means of solar panels, it could be pre-programmed to 'like' sunny places. That is, sunbathing would be a sub-goal. We would not need to give the robot any insight into why it liked these places, but other aspects of its behaviour would be learned and would adapt to the prevailing circumstances by trial and error. The pre­programmed 'chunks' of behaviour could be used along with learned 'chunks' to form the more complex behavioural patterns which would help it to survive.

If someone unaware of how the robot had been programmed was able to witness the robot's behaviour, they could be forgiven for thinking that the robot was driven by 'desire' for certain things. And such a person, who wanted to predict the behaviour of the robot, would be well advised to think in terms of 'desires' even if they knew about the internal program involved. The program would, after all, be extremely complex, and its precise behaviour would depend upon remembered data which was internal to the robot and was unknown to the observer. Trying to work things out from first principles would not be an efficient way of predicting the robot's behaviour in real-time.

The final touch, which would give the robot an almost human-like behaviour pattern, would be the ability to predict the behaviour of other organisms (i.e. people) based upon a similar model of their assumed 'desires'. Such a robot may well decide that helping others to achieve their 'desires' was the best way of achieving its own 'desires' (i.e. helping others was a sub-goal) and would, therefore, develop an analogue of altruistic behaviour.

We do not want to push this idea too far and suggest that this is actually how humans behave. It is merely a way of representing something which would otherwise be extremely elusive. Our robot may feel no inner sensations such as we feel, but its behaviour would make it look as if it did. It may seek its goals as an automaton without any awareness of a desire, but in our representation of its behaviour, our programs will identify goal states and will 'know' that the robot will try to achieve them. So 'desire' will not be repr sented by an actual feeling, but will instead become an inferred property of the programs which process the representations. We may not be able to create feelings, but we can write such programs.

This way of representing motivation is fundamental to the whole strategy we will develop in the next chapters. 'Goal' and 'anti-goal' states will be primitives of our representational scheme, and it is our contention that this simple mechanism is capable of supporting a representation of quite involved human social behaviour patterns such as those associated with the concepts 'duty', 'ownership', 'honesty' and 'flattery'.

21.2 Examples of Concepts involving Motivation

Many words contain hidden references to the mental states and motivations of people. We will try to illustrate this with a series of examples, and show how these representations can be constructed using 'goal' and 'anti-goal' states as primitives. Since the representations can become quite baroque, we will simplify them by showing only the relevant items. The notation {Sn .... } is used to indicate a state. It is identified by its number 'Sn' where 'n' is an integer. A causal connection between SI and S2 will be denoted SI->S2.

Example 1: likes

The representation of 'likes', as in 'John likes X', must identify X as the cause of John's potential happiness (in John's mind). Thus we have:

'John  likes  X'
owner  =  {John}
{SI  agent}
{S2  X}
{S3  goal  of  agent} 
{S4  S2->S3}



S2 and S4 are labelled 'potential', interpreted as 'John thinks that X causes him to achieve goal-state' It will be noted that this representation cannot distinguish between 'likes', 'desires' and similar words. A more adequate representation would make use of the representation of X. For example, if X = 'apples' then the representation of X would contain a description of X as something which people eat. This could be used to explain more fully how X would cause John to achieve his goal state by unifying John with the agent of the eating process. This is of course a very selfISh form of liking.



Example 2: loves

The idea is that 'loves' is the unselfish counterpart of 'likes', i.e. 'l fX loves Y then X wants to give pleasure to Y".

'X  loves  Y' 
owner  =  X
{Sl  X} 
{S2  Y}
{S3  goal  of  Y} 
{S4  goal  of  X}
  {S5  S3->S4}



Or, 'Y being happy makes X happy'.

Example 3: controls

The idea behind 'controls' is that one entity, X, can manipulate another, Y, in order to achieve the desired results of X.

'X  controls  Y'



owner  = 



{Sl  X} 
{S2  Y} 
{S3  some  arbitrary  state} 
{S4  goal  of  X}
{S5  S3->S4} 
{S6  S2->S3} 
{S7  SI->S6}



Note the last element. It indicates that X causes Y to be the cause of the state causing X's goal state.

Example 4: accepts

The interpretation of 'accepts' here is concerned with the acceptance of a situation or a state of affairs, not the acceptance of a gift. That would require a different representation.

'X  accepts  Y'



owner  = 



"{S1  X} 
{S2  Y}
{S3  an  arbitrary  state  of  affairs} 
{S4  goal  of  X}
{S5  not(S3->S4}}
{S6  S2->S3}
{S7  not  S1  controls  S2}



The idea here is that Y is causing a state S3 which is not causing X to achieve his goal state. Nevertheless, X does not seek to change this state of affairs (does not control Y).

A more subtle representation might include information about whether or not X was able to control Y if he wished (possible or not possible X control Y).

Example 5: owns

The word 'owns' is particularly difficult to represent because of the many different types of ownership which are possible (see section 16.2). The idea developed here is that ownership is a kind of social contract between the owner of something and the rest of humanity which accepts the owner's right to control the object 'owned'. This contract can be time limited so that it only lasts for the duration of a meeting (owning a chair) or for a lifetime (owning a leg). The acceptance by humanity may be limited to certain types of control. For example, parents 'own' their children but are not granted the power of life and death over them by society.

'X  owns  Y'



owner  =



{S1  X} 
{S2  Y}
{S3  humanity}
{S4  X  controls  Y} 
{S5  S3  accepts  S4}



For a representation of humanity we would require some of the ideas developed in Chapter 26.

Example 6: duty

Some will object strongly to the ideas we put forward here because they present a very poor view of human nature. The idea is that duty is a person's belief that certain behaviour is expected of them by society, and that should they fail to behave in this way society would disapprove. A more subtle representation might include the idea that the person concerned is not consciously aware of the details of this situation (or has forgotten) and is now aware only of a compulsion to serve humanity. He now seeks to please himself by this behaviour rather than humanity.

'X  has  a  duty  Y'



owner  =  X



{SI  X} 
{S2  Y}
{S3  humanity}
{S4  an  arbitrary  state}
{S5  S1->S4}  (S5  is  identified  with  'Y') 
{S6  goal  of  S3}
{S7  S4->S6}
{SS  S3  loves  S1}
{S9  S5->S8}
{S10  goal  of  X}
{S11  S8->S10}



Or, X takes action to bring about an arbitrary state which will cause the goal state of humanity to be achieved. The fact that X caused this state of affairs causes humanity to 'love' X. The goal of X is to have humanity loving him. If we dropped some of the intermediate causal links so that the achievement of humanity's goal was an end in itself for X (without the need for humanity to love him), we would have a less base interpretation of 'duty'.

Example 7: deceives

The idea here is that X causes Y to have a false idea about the true state of affairs (or at least what X believes to be true).

'X  deceives  Y  about  Z'



owner  =  X



{Sl  X} 
{S2  Y}
{S3  an  arbitrary  state  Z} 
{S4          owner  =  Y          {S5  not(S3)  } 
{S6S1->S4}



In other words, X causes Y to have a model (S4) in which the state (S5) is the opposite of the actual state of affairs (S 3). A more complex defmition might have X causing Y to believe that the achievement of some state will help him (Y) to achieve his goal, when it will actually help X to achieve his goal.

Example 8: ability

The idea is that X is able to cause certain things.

'X  has  the  ability  Y  to  Z'



owner  =



{S1  X}
{S2  arbitrary  state  or  set  of  states} 
{S3  S1->S2  'Y'}



Y is identified as the causal link between X and some arbitrary state Z. That is, X has the ability to bring about Z.

Example 9: flattery

'X  flatters  Y'



id  =  M1 
owner  =  me



{S1  X} 
{S2  Y} 
{S3      id  =  M2  owner  =  Y  {S4  id  =  M3  owner  =  X  {S5  Y  has  the  ability  Z}  {S6  X  likes  Z}}  (end  of  S4)}  (end  of  S3) 
{S5  Sl->S3}



For clarification we have introduced identifiers for each representation (Ml, M2 and M3). Y has a model (M2) in which he believes that X has a model (M3) in which X believes that Y has certain abilities, which X would like to have. That Y believes model (M2) was caused by X. Whether or not M2 is accurate is not shown here. Perhaps it should include 'X deceives Y about S5'.

With this we feel that we have plumbed the depths of the baser human attributes and leave the reader to try developing representations for such words as 'honesty', 'courage', 'trustworthiness' and 'intelligence'. It is fully accepted that these representations are simplistic travesties. What they do show, however, is that the representation of the ideas conveyed by these words is not entirely beyond the bounds of practicality. It also shows that complex defInitions can be constructed from simpler ones, holding open the prospect that we may be able to teach a computer the meaning of such terms in words which it already understands.