PRINCIPLES AND ARCHITECTURES IN AUTONOMOUS CONTROL AN INTRODUCTION

1
PRINCIPLES AND ARCHITECTURES IN AUTONOMOUS
CONTROL AN INTRODUCTION

• Sandor M Veres
• University of Southampton

2
Outline of the talk

• Autonomous intelligent systems
• Agents for autonomous controlfundamentals
• Architectures of agents
• Software engineering perspective
• The future?

3
Autonomous intelligent control systems

• Main initial areas of applications
• Operations of large utility networks
  (communication networks and Internet
  applications, water distribution and sewage,
  power generation and distribution, etc.)
• Automated manufacturing
• Unmanned vehicles (spacecrafts, UAVs, AUVs,
  autonomous ground vehicles, etc. )
• Robots and household appliances ( gardening and
  leisure boats, etc. )

4
What is intelligence?

• Intelligence is the ability of a system to
  act appropriately in an uncertain environment,
  where an appropriate action is that which
  increases the probability of success, and success
  is the achievement of behavioural sub-goals that
  support the systems ultimate goal.
• Alexander M. Yestel and James S. Albus
• National Institute of Standards and Technology
  (USA)

5
Agents for autonomous control the fundamentals

• Agents are autonomous computational entities
  that can be viewed as perceiving their
  environment through sensors and acting upon their
  environment through effectors intelligent
  indicates that the agents pursue their goals and
  execute their tasks such that they optimise some
  given performance measures. To say that agents
  are intelligent does not mean that they are
  omniscient or omnipotent, nor does it mean that
  they never fail. Rather, it means that they
  operate flexibly and rationally in a variety of
  environmental circumstances , given the
  information they have and their perceptual and
  effectual abilities..
• G. Weiss

6
Standard agents definition

• the agents environment S s1, s2,
• action set Aa1,
  a2,
• standard agent is defined by
• 
  action S ? A
• (maps sequences of environment states to
  actions)
• environment model env S ? A ? ?(S)
• reactive agent action S ? A

7
Formalization of agent behaviour

• A history of the agent/environment interaction
  is a sequence
• h s0 ?a1 ? s1 ?a2 ? s2 ?a3 ? s3 ?a4
  ? s4
• The characteristic behaviour of an agent action
  S ? A in an environment env S ? A ? ?(S)
  is the set of all possible histories
• If some property ? holds for all these
  histories, this property is called an invariant
  property of the agent and the environment.
• Two agents are behaviourally equivalent with
  respect to an environment, if their set of all
  histories are equal with respect to that
  environment.
• They are also called simply behaviourally
  equivalent if they are equivalent with respect to
  any environment.

8
Filling in more details

• The first architectural issue is to split the
  action mapping into perception and action
• see S ? P ,
  action P ? A
• where P is a nonempty set of perceptions.
• Agent states K . The next state is defined by
  a function
• next K ? P ?
  K
• and the action is decided by a new function
• action K ?
  A

9
ARCHITECTURES OF AGENTS

• logic based agents
• reactive agents
• belief-desire-intention agents
• layered architectures

10
Logic based architectures

• Let L be a set of sentences of classical
  first-order logic, and let D?(L) be the set of
  L databases, i.e. the set of sets of L-formulae.
• The internal state of an agent is then an element
  of D.
• An agents decision making process is modelled
  through a set of deduction rules ? . Write ? ?
  ? if the formula ? can be proved from the
  database ? using only the deduction rules ?.
• The logic-based agents perception function see
  remains unchanged see S ? P
• The next function has the form
• next D ? P ? D
• which maps a database and a percept to a
  new database.

11
Logic based architectures (page 2)

• The agents action selection function
• action D
  ? A
• is defined in terms of its deduction rules. This
  assumes a mapping Do A ? L that associates a
  formula with each action. For each ? ? D the
  following procedure is followed to obtain
  action(?)
• For each action a ? A, it is tested whether ?
  ? Do(a) , the first a that satisfies this will
  be action(?) .
• If no such a is found then by going through all
  A, then it is examined for all a ? A, whether for
  some a ? / ? ?Do(a), which means that the
  negation of Do(a) cannot be derived. The first a
  that satisfies this will become action(?). This
  ensures that at least a logically consistent
  action will be selected by the agent.

12
Logic based architectures (page 3)

• Calculative rationality is not acceptable in
  environments that change faster than the agent
  can make decisions.
• Another potential problem is that representing
  properties of some dynamic, real-world
  environments is sometimes extremely hard by
  logic.
• A good number of pure logical approaches have
  been developed to agents since the early 80s and
  throughout the 90s, for instance well know
  systems are METATEM by Michael Fisher and
  CONGOLOG by Y Lésperance et al.

13
Reactive architectures

• The problems with the computational
  complexity of the logic based architectures
  inspired the development of reactive
  architectures. The term reactive is used because
  such agents are often perceived as simply
  reacting to the environment. The best known is
  the subsumption architecture.
• The function see that defines the agents
  perceptual ability is the same as before
• see S ? P

14
Reactive architectures (page 2)

• A behaviour is a pair (c,a) where c ? P is a set
  of perceptions, also called the condition, and a
  ? A is an action.
• The agents set of behaviour rules will be
  defined by a set B of behaviours .
• One says that a behaviour fires when the
  environment is in state s ? S iff see(s) ? c .
• A binary inhibition relation ? ? B ? B is defined
  as a transitive, irreflexive and antisymmetric
  relation.

15
Reactive architectures (page 3)

• The inhibition relation ? is what is commonly
  referred to as the subsumption hierarchy, which
  is defined over the set of behaviours B.
• Definition of a action(p)
• when a behaviour (c,a) is found that is not
  inhibited in any way, then that action defines
  a action(p) .

16
Reactive architectures (page 4)

• Advantages
• The overall time complexity is not greater
  than O(nn) where n is the greater of the number
  of behaviours or perceptions. This allows
  strictly bounded and small decision time in
  realtime applications. Also subsumption is
  elegant.
• Problems
• They do not employ models of their
  environment, they must have sufficient
  information available to them in their local
  environment to determine their actions. Difficult
  to see how purely reactive agents can be made to
  learn from their experience.

17
Literature (reactive agents)

• Subsumption architectures by R Brooks.
• The agent network architecture by P Maes.
• The work of Rosenschein and Kaelbling on situated
  automata they provide a method to compile
  agents specified in a logical framework into
  computationally efficient very simple machines.

18
Beliefs-desire-intention (BDI) architectures

• Practical reasoning of humans involves two
  important phases
• (1) what goals do we want to achieve ?
• (deliberation process )
• (2) how awe are going to achieve them ?
• (means-end reasoning)

19
Beliefs-desire-intention (BDI) architectures
(page 2)

• Let Bel, Des and Int denote large abstract sets
  from which beliefs, desires and intentions can
  be taken.
• The state of a BDI agent is at any moment a
  triple (B,D,I) where B?Bel, D?Des and I ? Int .
  
• An agents belief revision function (brf) is a
  mapping from a belief set and percept into a new
  belief set
• brf ?(Bel) ? P ?
  ?(Bel)
• This updates the agents belief set based on
  the agents perception of its environment.

20
BDI architectures (desire generation)

• The options generation function (options) maps a
  set of beliefs and a set of intentions to a set
  of desires
• options ?(Bel) ? ?(Int) ? ?(Des)
• The main function of options is means-end
  reasoning, that must be consistent with beliefs
  and current intentions and also opportunistic to
  recognise when environmental circumstances change
  advantageously.

21
BDI architectures (deliberation process)

• The deliberation process, i.e. deciding what to
  do, is represented by the filter function
• filter ?(Bel) ? ?(Des) ? ?(Int) ? ?(Int)
  
• It must drop intentions that are no longer
  achievable, retain intentions that are not yet
  achieved and it should adopt new intentions to
  achieve existing intentions or to exploit new
  opportunities.

22
BDI architectures (deliberation process)

• Finally, the function execute is used to select
  an executable intention, one that corresponds to
  a directly executable action
• execute
  ?(Int) ? A
• As in the standard agent architecture, the
  current behaviour is determined by the function
• action
  P ? A
• Given a p?P , the p ? a map of action is
  defined by
• Bbrf(B,p)
• Doptions(D,I)
• Ifilter(B,D,I)
• aexecute(I)

23
BDI architectures (block diagram literature)

• The procedural reasoning system (PRS) by M P
  Georgeff and Lansky
• A lot of work has been carried out to formalise
  the BDI model, see the ref. paper by A S Rao
• A BDI agent programming language is for instance
  JACK in Java.

sensor input
brf
Beliefs B
options
Means-end reasoning
Desires D
filter
deliberation
Intentions I
execute
Select and perform action
24
Layered architectures

• Vertical and horizontal layers of blocks for
  reactive, logic based and coordinated behaviours.
• US National Institute of Standards describes
  layered architectures for
• - Vehicle domain 4D/RC
• - Manufacturing Domain ISAM
• - Space Domain - NASREM
• Documents can be found at
• http//www.isd.mel.nist.gov/projects/rcs/referenc
  e.html .

25
Software engineering perspective

• AOP agent oriented programming programming
  agents in terms of mentalistic terms such as
  beliefs, desires and and to enable them to
  communicate with each other in terms of some
  ontology.
• List of languages at http//www.dsv.su.se/mab/AOP
  /AgentLinks.html

26
Future?

• This area, when applied to engineering, will
  probably bring the most significant technological
  developments to change our lives this century.
• Opportunities 8
• ----------------------
  --
• Thank you for your attention!