This presentation introduces the fundamentals of contemporary AI research and highlights a significant challenge that we have still not addressed - namely that we have to trade quality of decision making against speed of decision making. It goes on to discuss the concepts behind the "Integrated Influence Architecture", a new approach to making high-speed and high-quality decisions currently under development at University of Strathclyde.

1. Integrated Influence :
The Six Million Dollar Man
of AI
L Dicken

2. What is AI?
2

3. What is AI?
• Any time a computer makes any sort of decision
between a number of options, it can be thought of
as acting “intelligently”.
2

4. What is AI?
• Any time a computer makes any sort of decision
between a number of options, it can be thought of
as acting “intelligently”.
• Whether or not those decisions are the right ones
is how “good” the intelligence is.
2

5. AI Applications
3

6. AI Applications
• Automatic Translation
3

7. AI Applications
• Automatic Translation
• Statistical Analysis
3

8. AI Applications
• Automatic Translation
• Statistical Analysis
• Optimising Resource Usage
3

9. AI Applications
• Automatic Translation
• Statistical Analysis
• Optimising Resource Usage
• Scheduling Problems
3

10. AI Applications
• Automatic Translation
• Statistical Analysis
• Optimising Resource Usage
• Scheduling Problems
• Automated Planning
3

11. AI Applications
• Automatic Translation
• Statistical Analysis
• Optimising Resource Usage
• Scheduling Problems
• Automated Planning
• Image/Facial Recognition
3

12. AI Applications
• Automatic Translation
• Statistical Analysis
• Optimising Resource Usage
• Scheduling Problems
• Automated Planning
• Image/Facial Recognition
• And many more...
3

13. Basics
4

14. Basics
• Broadly, two conceptual paradigms in AI
‣ Reaction
‣ Deliberation
4

15. Basics
• Broadly, two conceptual paradigms in AI
‣ Reaction
‣ Deliberation
• Reaction aims to program “instinctive” reactions to
minimal subsets of stimuli.
4

16. Basics
• Broadly, two conceptual paradigms in AI
‣ Reaction
‣ Deliberation
• Reaction aims to program “instinctive” reactions to
minimal subsets of stimuli.
• Deliberation describes reasoning-based approaches,
using all the information available.
4

17. Automated Planning
5

18. Automated Planning
• AP is a deliberative technique
5

19. Automated Planning
• AP is a deliberative technique
• Given a description of
5

20. Automated Planning
• AP is a deliberative technique
• Given a description of
‣ Current state of the world
5

21. Automated Planning
• AP is a deliberative technique
• Given a description of
‣ Current state of the world
‣ Actions that can be applied and the way they affect the
world
5

22. Automated Planning
• AP is a deliberative technique
• Given a description of
‣ Current state of the world
‣ Actions that can be applied and the way they affect the
world
‣ A set of goals to be achieved
5

23. Automated Planning
• AP is a deliberative technique
• Given a description of
‣ Current state of the world
‣ Actions that can be applied and the way they affect the
world
‣ A set of goals to be achieved
• Automatically determines a sequence of actions that
will complete the task.
5

24. PDDL
6

25. PDDL
• Planning Domain Description Language
6

26. PDDL
• Planning Domain Description Language
• Propositional representation of world
6

27. PDDL
• Planning Domain Description Language
• Propositional representation of world
• All things not asserted true are false
6

28. PDDL
• Planning Domain Description Language
• Propositional representation of world
• All things not asserted true are false
• All things true now will remain true unless negated
6

29. PDDL
• Planning Domain Description Language
• Propositional representation of world
• All things not asserted true are false
• All things true now will remain true unless negated
• Extensions deal with a variety of extras
6

30. PDDL
• Planning Domain Description Language
• Propositional representation of world
• All things not asserted true are false
• All things true now will remain true unless negated
• Extensions deal with a variety of extras
‣ e.g. Numerical values, Temporal actions, Continuous
effects etc.
6

31. Example Domain
(define (domain logistics)
(:requirements :strips :typing)
(:types truck - vehicle
package vehicle - physobj
location - place
place physobj - object)
(:predicates (in-city ?loc - place ?city - city)
(at ?obj - physobj ?loc - place)
(in ?pkg - package ?veh - vehicle))
(:action LOAD
:parameters (?pkg - package ?truck - truck ?loc - place)
:precondition (and (at ?truck ?loc) (at ?pkg ?loc))
:effect (and (not (at ?pkg ?loc)) (in ?pkg ?truck)))
(:action UNLOAD
:parameters (?pkg - package ?truck - truck ?loc - place)
:precondition (and (at ?truck ?loc) (in ?pkg ?truck))
:effect (and (not (in ?pkg ?truck)) (at ?pkg ?loc)))
(:action MOVE
:parameters (?truck - truck ?loc-from - place ?loc-to - place ?city - city)
:precondition
(and (at ?truck ?loc-from) (in-city ?loc-from ?city) (in-city ?loc-to ?city))
:effect
(and (not (at ?truck ?loc-from)) (at ?truck ?loc-to)))
7

32. Planning Graphs

33. Planning Graphs
At(T1, L1)
At(P1, L1)
f0

34. Planning Graphs
Move(T1,
L2)
At(T1, L1)
Move(T1,
L3)
At(P1, L1)
Load(P
1,T1)
f0 a1

35. Planning Graphs
Move(T1,
L2)
At(T1, L1) At(T1, L1)
Move(T1,
L3)
At(P1, L1) In(P1, T1)
Load(P
1,T1)
f0 a1 f1

36. Planning Graphs
Move(T1, Move(T
L2) 1, L2)
At(T1, L1) At(T1, L1)
Move(T1, Move(T1,
L3) L3)
At(P1, L1) In(P1, T1)
Load(P Unload(P
1,T1) 1,T1)
f0 a1 f1 a2

37. Planning Graphs
Move(T1, Move(T
L2) 1, L2)
At(T1, L1) At(T1, L1) At(T1, L2)
Move(T1, Move(T1,
L3) L3)
At(P1, L1) In(P1, T1) In(P1, T1)
Load(P Unload(P
1,T1) 1,T1)
f0 a1 f1 a2 f2

38. Planning Graphs
Move(T1, Move(T Move(T1,
L2) 1, L2) L1)
At(T1, L1) At(T1, L1) At(T1, L2)
Move(T1, Move(T1, Move(T1,
L3) L3) L3)
At(P1, L1) In(P1, T1) In(P1, T1)
Load(P Unload(P Unload(
1,T1) 1,T1) P1,T1)
f0 a1 f1 a2 f2 a3

39. Planning Graphs
Move(T1, Move(T Move(T1,
L2) 1, L2) L1)
At(T1, L1) At(T1, L1) At(T1, L2) At(T1, L2)
Move(T1, Move(T1, Move(T1,
L3) L3) L3)
At(P1, L1) In(P1, T1) In(P1, T1) At(P1, L2)
Load(P Unload(P Unload(
1,T1) 1,T1) P1,T1)
f0 a1 f1 a2 f2 a3 f3

40. The Core of Planning
9

41. The Core of Planning
• Previous example made it look so easy.
9

42. The Core of Planning
• Previous example made it look so easy.
• Trivial example, worked out in advance.
9

43. The Core of Planning
• Previous example made it look so easy.
• Trivial example, worked out in advance.
• At every action layer, many choices don’t help
9

44. The Core of Planning
• Previous example made it look so easy.
• Trivial example, worked out in advance.
• At every action layer, many choices don’t help
‣ Easy to disappear down a rabbit hole
9

45. The Core of Planning
• Previous example made it look so easy.
• Trivial example, worked out in advance.
• At every action layer, many choices don’t help
‣ Easy to disappear down a rabbit hole
• Planning all about guiding search across the action/
fact layer space.
9

46. The Core of Planning
• Previous example made it look so easy.
• Trivial example, worked out in advance.
• At every action layer, many choices don’t help
‣ Easy to disappear down a rabbit hole
• Planning all about guiding search across the action/
fact layer space.
‣ Different heuristics, search strategies, pruning techniques
9

47. Problems
10

48. Problems
• Search space is massive.
10

49. Problems
• Search space is massive.
‣ Computational complexity high
10

50. Problems
• Search space is massive.
‣ Computational complexity high
‣ Processing time required also high
10

51. Problems
• Search space is massive.
‣ Computational complexity high
‣ Processing time required also high
• Not only that but models used are “abstractions”
10

52. Problems
• Search space is massive.
‣ Computational complexity high
‣ Processing time required also high
• Not only that but models used are “abstractions”
‣ Typically removes chance to fail an action
10

53. Problems
• Search space is massive.
‣ Computational complexity high
‣ Processing time required also high
• Not only that but models used are “abstractions”
‣ Typically removes chance to fail an action
‣ Typically removes other agents and the consequences of
their actions
10

54. Problems
• Search space is massive.
‣ Computational complexity high
‣ Processing time required also high
• Not only that but models used are “abstractions”
‣ Typically removes chance to fail an action
‣ Typically removes other agents and the consequences of
their actions
‣ Typically removes a lot of the detail e.g. Driver for truck
10

55. Time Constraints
11

56. Time Constraints
• International Planning Competition entrants get
around 30m to generate a plan for a single problem.
11

57. Time Constraints
• International Planning Competition entrants get
around 30m to generate a plan for a single problem.
• AAAI General Game Playing entrants get 5-10s to
decide on their next move.
11

58. Time Constraints
• International Planning Competition entrants get
around 30m to generate a plan for a single problem.
• AAAI General Game Playing entrants get 5-10s to
decide on their next move.
• Games Industry aims for 60fps execution - 16ms
per frame.
11

59. Time Constraints
• International Planning Competition entrants get
around 30m to generate a plan for a single problem.
• AAAI General Game Playing entrants get 5-10s to
decide on their next move.
• Games Industry aims for 60fps execution - 16ms
per frame.
‣ Most of that is spent on graphics, physics etc.
11

60. Time Constraints
• International Planning Competition entrants get
around 30m to generate a plan for a single problem.
• AAAI General Game Playing entrants get 5-10s to
decide on their next move.
• Games Industry aims for 60fps execution - 16ms
per frame.
‣ Most of that is spent on graphics, physics etc.
‣ AI gets maybe 1ms to work out everything it needs to
11

61. Reactive AI
12

62. Reactive AI
• Reactive AI makes snap decisions based on current
state of the world.
12

63. Reactive AI
• Reactive AI makes snap decisions based on current
state of the world.
• More tolerant to action failure - one action isn’t
part of a long chain of actions that depend on it.
12

64. Reactive AI
• Reactive AI makes snap decisions based on current
state of the world.
• More tolerant to action failure - one action isn’t
part of a long chain of actions that depend on it.
• Typically gives a very good response time to input
received from the environment.
12

65. Subsumption Arch.
13

66. Subsumption Arch.
• Quintessential reactive approach.
13

67. Subsumption Arch.
• Quintessential reactive approach.
• Library of behaviours ordered by priority.
13

68. Subsumption Arch.
• Quintessential reactive approach.
• Library of behaviours ordered by priority.
• Each behaviour maps detected input to relevant
response.
13

69. Subsumption Arch.
• Quintessential reactive approach.
• Library of behaviours ordered by priority.
• Each behaviour maps detected input to relevant
response.
• Higher priority behaviours are able to “subsume” or
override the output of the lower priority ones.
13

70. Subsumption
Behaviour 1
Behaviour 2
Input Mux
Behaviour 3
Behaviour 4
Output
14

71. Influence Maps
15

72. Influence Maps
• Much more simplistic approach
15

73. Influence Maps
• Much more simplistic approach
• Influence radiates from objects similarly to magnetic
fields.
15

74. Influence Maps
• Much more simplistic approach
• Influence radiates from objects similarly to magnetic
fields.
• Good things attract the agent, bad things repel the
agent.
15

75. Influence Maps
• Much more simplistic approach
• Influence radiates from objects similarly to magnetic
fields.
• Good things attract the agent, bad things repel the
agent.
• Interaction of influence is typically (but not
necessarily) additive.
15

76. Influence Maps
16

77. Neural Networks
17

78. Stateful vs. Stateless
18

79. Stateful vs. Stateless
• Deliberative reasoning is by its nature stateful
18

80. Stateful vs. Stateless
• Deliberative reasoning is by its nature stateful
• Reactive systems typically are stateless
18

81. Stateful vs. Stateless
• Deliberative reasoning is by its nature stateful
• Reactive systems typically are stateless
• Trying to retrofit them to include states typically
adds the type of complexity they were designed to
avoid.
18

82. Stateful vs. Stateless
• Deliberative reasoning is by its nature stateful
• Reactive systems typically are stateless
• Trying to retrofit them to include states typically
adds the type of complexity they were designed to
avoid.
‣ E.g. Trying to capture state in a NN involves a separate
NN designed to have a delayed feedback into the input
18

83. Stateful vs. Stateless
• Deliberative reasoning is by its nature stateful
• Reactive systems typically are stateless
• Trying to retrofit them to include states typically
adds the type of complexity they were designed to
avoid.
‣ E.g. Trying to capture state in a NN involves a separate
NN designed to have a delayed feedback into the input
• Reactive Systems struggle to make long term plans
18

84. Limitations of AI
19

85. Limitations of AI
• Contemporary AI is capable of very sophisticated
insightful decision making.
‣ ....eventually
• Also able to make very rapid decision making.
‣ ...at the expense of long-term decision quality
• Range of problems require both high quality and
short time frame decisions.
19

86. Six Million Dollar Man
20

87. Six Million Dollar Man
• We can rebuild it!
20

88. Six Million Dollar Man
• We can rebuild it!
‣ Better
20

89. Six Million Dollar Man
• We can rebuild it!
‣ Better
‣ Stronger
20

90. Six Million Dollar Man
• We can rebuild it!
‣ Better
‣ Stronger
‣ Faster
20

91. Six Million Dollar Man
• We can rebuild it!
‣ Better
‣ Stronger
‣ Faster
20

92. Six Million Dollar Man
• We can rebuild it!
‣ Better
‣ Stronger
‣ Faster
• But, we don’t yet have
the technology...
20

93. Integrated Influence
21

94. Integrated Influence
• My work focuses on trying to bridge the gap
between reaction and deliberation in novel ways
21

95. Integrated Influence
• My work focuses on trying to bridge the gap
between reaction and deliberation in novel ways
• Previous approaches have typically either :
21

96. Integrated Influence
• My work focuses on trying to bridge the gap
between reaction and deliberation in novel ways
• Previous approaches have typically either :
‣ Created an agent that deliberates about certain aspects
of the world and reacts to others
21

97. Integrated Influence
• My work focuses on trying to bridge the gap
between reaction and deliberation in novel ways
• Previous approaches have typically either :
‣ Created an agent that deliberates about certain aspects
of the world and reacts to others
‣ Created an agent that reacts within the parameters of a
deiberatively generated trajectory
21

98. Integrated Influence
• My work focuses on trying to bridge the gap
between reaction and deliberation in novel ways
• Previous approaches have typically either :
‣ Created an agent that deliberates about certain aspects
of the world and reacts to others
‣ Created an agent that reacts within the parameters of a
deiberatively generated trajectory
• Neither approach has proven particularly robust
21

99. Concept
22

100. Concept
• We take the view that many aspects of the world
can’t be tackled by one or other paradigm, but
require both.
22

101. Concept
• We take the view that many aspects of the world
can’t be tackled by one or other paradigm, but
require both.
• To this end, our architecture aims to constantly use
all information available, both deliberative and
reactive, to make decisions
22

102. Search vs. Evaluation
23

103. Search vs. Evaluation
• Searching spaces is a complex task.
‣ Typically at least NP-Hard, PDDL domains can be as
complex as PSPACE-Complete
• What if, instead of performing search, we could
reformulate the problem into something closer to
function evaluation?
23

104. Propositions
24

105. Propositions
• PDDL’s propositional representation gives a state
representation of very high dimension, with each
dimension having exactly two possible values.
• Can we do better with another representation
format?
24

106. SAS+
25

107. SAS+
• SAS+ groups mutually exclusive PDDL props
together.
‣ Propositional - at(P1, L1), at(P1, L2), at(P1, L3), in(P1, T1)
‣ SAS+ - locationP1 ∈ {L1, L2, L3, T1}
25

108. SAS+
• SAS+ groups mutually exclusive PDDL props
together.
‣ Propositional - at(P1, L1), at(P1, L2), at(P1, L3), in(P1, T1)
‣ SAS+ - locationP1 ∈ {L1, L2, L3, T1}
• Also captures the ordering that the values take
‣ E.g. From any Lx to Ly, locationP1 take value T1 between
25

109. SAS+
• SAS+ groups mutually exclusive PDDL props
together.
‣ Propositional - at(P1, L1), at(P1, L2), at(P1, L3), in(P1, T1)
‣ SAS+ - locationP1 ∈ {L1, L2, L3, T1}
• Also captures the ordering that the values take
‣ E.g. From any Lx to Ly, locationP1 take value T1 between
• Identifies the dependencies between different types
of object
25

110. DTGs
T1
L1 L2 L3
26

111. Influence Landscapes
• Introduced the concept of Influence Map earlier
27

112. Influence Landscapes
• Introduced the concept of Influence Map earlier
• Influence Landscapes extend the idea away from a
purely spatial representation and into a conceptual
representation.
27

113. Influence Landscapes
• Introduced the concept of Influence Map earlier
• Influence Landscapes extend the idea away from a
purely spatial representation and into a conceptual
representation.
• Allows for the same function-based approach to be
applied to reasoning.
27

114. Influence Landscape
T1
L1 L2 L3
28

115. Caveats
29

116. Caveats
• It isn’t quite this easy
29

117. Caveats
• It isn’t quite this easy
• Need to consider ‘Causal Links’
29

118. Caveats
• It isn’t quite this easy
• Need to consider ‘Causal Links’
• To load the package at L1, the truck needs to be at
L1 too.
29

119. Caveats
• It isn’t quite this easy
• Need to consider ‘Causal Links’
• To load the package at L1, the truck needs to be at
L1 too.
• Interlinked set of DTGs allow this to be captured
and represented.
29

120. DTG with CG
T1
L1 L2 L3
30

121. DTG with CG
T1
L1 L2 L3
L1 L2 L3
30

122. Landscape Generators
31

123. Landscape Generators
• Previous example generated using a simple critical
path analysis.
31

124. Landscape Generators
• Previous example generated using a simple critical
path analysis.
• Need to get much more informed view of the
world around the agent.
31

125. Landscape Generators
• Previous example generated using a simple critical
path analysis.
• Need to get much more informed view of the
world around the agent.
• Illustrates the concept though - also useful for
providing the “naive” view of the world.
31

126. Landscape Generators
• Previous example generated using a simple critical
path analysis.
• Need to get much more informed view of the
world around the agent.
• Illustrates the concept though - also useful for
providing the “naive” view of the world.
• Critical path analysis gives some info about the
structure of the world, but is not fully informed.
31

127. Stacks
32

128. Stacks
• Stacks are the name given to Landscape Generators
32

129. Stacks
• Stacks are the name given to Landscape Generators
• Each stack is tasked with assigning a numerical value
to every node within the DTG/CG space.
32

130. Stacks
• Stacks are the name given to Landscape Generators
• Each stack is tasked with assigning a numerical value
to every node within the DTG/CG space.
‣ Remember that this is a smaller space than propositional
32

131. Stacks
• Stacks are the name given to Landscape Generators
• Each stack is tasked with assigning a numerical value
to every node within the DTG/CG space.
‣ Remember that this is a smaller space than propositional
• Each stack deals with a specific aspect of the world
or a specific approach
32

132. Stacks
• Stacks are the name given to Landscape Generators
• Each stack is tasked with assigning a numerical value
to every node within the DTG/CG space.
‣ Remember that this is a smaller space than propositional
• Each stack deals with a specific aspect of the world
or a specific approach
‣ E.g. Reactive, Deliberative (or some other information)
32

133. Deliberative - Plan
33

134. Deliberative - Plan
• We can guide the search by bringing in the
information a deliberative reasoner provides.
33

135. Deliberative - Plan
• We can guide the search by bringing in the
information a deliberative reasoner provides.
‣ E.g. Automated Planning
33

136. Deliberative - Plan
• We can guide the search by bringing in the
information a deliberative reasoner provides.
‣ E.g. Automated Planning
• We can implement this as a stack generating a
landscape reflecting the influence the plan exerts on
our agent.
33

137. Deliberative - Plan
• We can guide the search by bringing in the
information a deliberative reasoner provides.
‣ E.g. Automated Planning
• We can implement this as a stack generating a
landscape reflecting the influence the plan exerts on
our agent.
• Typically this will be a best-case assumption.
33

138. Conformity
34

139. Conformity
• How heavily do we want to reward conforming to
the plan?
34

140. Tight Conformity
35

141. Tight Conformity
• Reward every node the plan requires the agent to
visit in the graph, Royal Road style.
35

142. Tight Conformity
• Reward every node the plan requires the agent to
visit in the graph, Royal Road style.
• Visualise landscape as a ridge to the summit
35

143. Tight Conformity
• Reward every node the plan requires the agent to
visit in the graph, Royal Road style.
• Visualise landscape as a ridge to the summit
‣ Excellent in best-cases
35

144. Tight Conformity
• Reward every node the plan requires the agent to
visit in the graph, Royal Road style.
• Visualise landscape as a ridge to the summit
‣ Excellent in best-cases
‣ Poor when flexibility required
35

145. Tight Conformity
• Reward every node the plan requires the agent to
visit in the graph, Royal Road style.
• Visualise landscape as a ridge to the summit
‣ Excellent in best-cases
‣ Poor when flexibility required
• After deviating from the plan, best approach seems
to be to rejoin the ridge - this may not be the case.
35

146. Loose Conformity
36

147. Loose Conformity
• Use the plan to mark out a general path without
strictly defining each node of the plan.
36

148. Loose Conformity
• Use the plan to mark out a general path without
strictly defining each node of the plan.
• Much more flexible approach, guides the agent
rather than dictating to it.
36

149. Loose Conformity
• Use the plan to mark out a general path without
strictly defining each node of the plan.
• Much more flexible approach, guides the agent
rather than dictating to it.
• But how do you determine which nodes to mark
and which to ignore?
36

150. Focal Nodes
37

151. Focal Nodes
• Focal Nodes are these waypoints in the plan.
37

152. Focal Nodes
• Focal Nodes are these waypoints in the plan.
• Previous work with SAS+ has shown that a DTG
can be deformed to be laid out in any way.
37

153. Focal Nodes
• Focal Nodes are these waypoints in the plan.
• Previous work with SAS+ has shown that a DTG
can be deformed to be laid out in any way.
‣ Logistics-style domain overlaid on a map of Europe.
37

154. Focal Nodes
• Focal Nodes are these waypoints in the plan.
• Previous work with SAS+ has shown that a DTG
can be deformed to be laid out in any way.
‣ Logistics-style domain overlaid on a map of Europe.
• Highlights by inspection clumps of nodes and
connections between them
37

155. Focal Nodes
• Focal Nodes are these waypoints in the plan.
• Previous work with SAS+ has shown that a DTG
can be deformed to be laid out in any way.
‣ Logistics-style domain overlaid on a map of Europe.
• Highlights by inspection clumps of nodes and
connections between them
‣ E.g. Channel Tunnel, Dover ferry etc.
37

156. Clustering
38

157. Clustering
• We can pick out the FNs by hand, but that’s no fun.
38

158. Clustering
• We can pick out the FNs by hand, but that’s no fun.
• Instead, using the structure of the graph to find
them automagically.
38

159. Clustering
• We can pick out the FNs by hand, but that’s no fun.
• Instead, using the structure of the graph to find
them automagically.
• Clustering the nodes of the graph allows us to
group nodes together by proximity
38

160. Clustering
• We can pick out the FNs by hand, but that’s no fun.
• Instead, using the structure of the graph to find
them automagically.
• Clustering the nodes of the graph allows us to
group nodes together by proximity
• Fuzzy Clustering allows us to identify nodes that lie
between groups
38

161. Fuzzy Clustering
39

162. Focal Nodes
40

163. Using Focal Nodes
41

164. Using Focal Nodes
• FNs can be identified offline for every DTG in the
domain.
41

165. Using Focal Nodes
• FNs can be identified offline for every DTG in the
domain.
• FNs that the plan indicates should be passed
through then become “Activated”.
41

166. Using Focal Nodes
• FNs can be identified offline for every DTG in the
domain.
• FNs that the plan indicates should be passed
through then become “Activated”.
• These nodes are given influence in the landscape
and this is propagated out across the graph to guide
the agent to these key nodes.
41

167. Plan Data
42

168. Environmental Data
43

169. Environmental Data
• We can overcome the deficiencies in deliberative
landscape by bringing in data about the environment
43

170. Environmental Data
• We can overcome the deficiencies in deliberative
landscape by bringing in data about the environment
• Gives the kind of insight that a reactive system
would use to make decisions.
43

171. Environmental Data
• We can overcome the deficiencies in deliberative
landscape by bringing in data about the environment
• Gives the kind of insight that a reactive system
would use to make decisions.
• Allows us to inform the agent about things that may
require it to deviate from the planned trajectory
43

172. Preferences
44

173. Preferences
• Preferences allow the agent to bias the influence of
nodes in the graph at execution time based on data
being sensed.
44

174. Preferences
• Preferences allow the agent to bias the influence of
nodes in the graph at execution time based on data
being sensed.
• Can be either positive or negative.
44

175. Preferences
• Preferences allow the agent to bias the influence of
nodes in the graph at execution time based on data
being sensed.
• Can be either positive or negative.
• Applies an influence of appropriate strength to the
target node and then propagates that out
44

176. Road Blocks
45

177. Road Blocks
• Road Block is an edge that should be in the domain
but for some reason is not traversable at this time.
45

178. Road Blocks
• Road Block is an edge that should be in the domain
but for some reason is not traversable at this time.
• Two conceptual models
45

179. Road Blocks
• Road Block is an edge that should be in the domain
but for some reason is not traversable at this time.
• Two conceptual models
‣ Cancelled flight - this edge will never be traversable
45

180. Road Blocks
• Road Block is an edge that should be in the domain
but for some reason is not traversable at this time.
• Two conceptual models
‣ Cancelled flight - this edge will never be traversable
‣ Blocked road - this edge may be traversable later
45

181. Road Blocks
• Road Block is an edge that should be in the domain
but for some reason is not traversable at this time.
• Two conceptual models
‣ Cancelled flight - this edge will never be traversable
‣ Blocked road - this edge may be traversable later
• Should the edge be removed?
45

182. Road Blocks
• Road Block is an edge that should be in the domain
but for some reason is not traversable at this time.
• Two conceptual models
‣ Cancelled flight - this edge will never be traversable
‣ Blocked road - this edge may be traversable later
• Should the edge be removed?
‣ Opted to implement the ‘Road Block’ model as this
allows resensing later to check the state of the edge.
45

183. Integrated Landscape
46

184. Integrated Landscape
• By combining the landscape from each of the
individual stacks, we get the “Integrated Influence
Landscape” from which the architecture draws its
name.
46

185. Integrated Landscape
• By combining the landscape from each of the
individual stacks, we get the “Integrated Influence
Landscape” from which the architecture draws its
name.
• Currently using an unweighted additive model for
combining
46

186. Integrated Landscape
• By combining the landscape from each of the
individual stacks, we get the “Integrated Influence
Landscape” from which the architecture draws its
name.
• Currently using an unweighted additive model for
combining
‣ This may prove to be sub-optimal in further testing
46

187. Using the IIL
47

188. Using the IIL
• Given an IIL, the agent can then climb the gradient
towards the goal, changing between DTGs as
required by the Causal Graph.
47

189. Using the IIL
• Given an IIL, the agent can then climb the gradient
towards the goal, changing between DTGs as
required by the Causal Graph.
• Hill Climbing algorithm obvious choice, but gets
stuck at local maxima
47

190. Using the IIL
• Given an IIL, the agent can then climb the gradient
towards the goal, changing between DTGs as
required by the Causal Graph.
• Hill Climbing algorithm obvious choice, but gets
stuck at local maxima
‣ Experimented with Forced-movement Hill Climbing
47

191. Using the IIL
• Given an IIL, the agent can then climb the gradient
towards the goal, changing between DTGs as
required by the Causal Graph.
• Hill Climbing algorithm obvious choice, but gets
stuck at local maxima
‣ Experimented with Forced-movement Hill Climbing
‣ Also Neighbourhood-bounded A*
47

192. Using the IIL
• Given an IIL, the agent can then climb the gradient
towards the goal, changing between DTGs as
required by the Causal Graph.
• Hill Climbing algorithm obvious choice, but gets
stuck at local maxima
‣ Experimented with Forced-movement Hill Climbing
‣ Also Neighbourhood-bounded A*
‣ Currently using Forced HC with ties broken randomly
47

193. Integrated Influence
Plan Landscape Env Landscape
Structure Landscape
48

194. Integrated Influence
Integrated Influence Landscape
49

195. Bonus Feature!
50

196. Bonus Feature!
• Brought up earlier time constraints of the Game
Industry - around 1ms per frame.
50

197. Bonus Feature!
• Brought up earlier time constraints of the Game
Industry - around 1ms per frame.
• Games give us a great context for testing “real
world” style applications, and often already have the
fast/smart requirements
50

198. Bonus Feature!
• Brought up earlier time constraints of the Game
Industry - around 1ms per frame.
• Games give us a great context for testing “real
world” style applications, and often already have the
fast/smart requirements
• Fully controllable simulation environment
50

199. Bonus Feature!
• Brought up earlier time constraints of the Game
Industry - around 1ms per frame.
• Games give us a great context for testing “real
world” style applications, and often already have the
fast/smart requirements
• Fully controllable simulation environment
• Very pretty demos
50

200. Parrallelisation
51

201. Parrallelisation
• By its nature, the stack paradigm is very flexible
51

202. Parrallelisation
• By its nature, the stack paradigm is very flexible
• Designed for each stack to be updated
asynchronously
51

203. Parrallelisation
• By its nature, the stack paradigm is very flexible
• Designed for each stack to be updated
asynchronously
• Very suitable to parallel execution
51

204. Parrallelisation
• By its nature, the stack paradigm is very flexible
• Designed for each stack to be updated
asynchronously
• Very suitable to parallel execution
‣ Increasingly a big factor in modern computing
51

205. Vector Operations
52

206. Vector Operations
• The vast majority of the maths mentioned can be
stated as vector and matrix operations.
52

207. Vector Operations
• The vast majority of the maths mentioned can be
stated as vector and matrix operations.
• Makes the whole architecture very suitable to
execution on Cell SPUs.
52

208. Vector Operations
• The vast majority of the maths mentioned can be
stated as vector and matrix operations.
• Makes the whole architecture very suitable to
execution on Cell SPUs.
‣ Synergistic Processing Units are vector-based
coprocessors in Cell-based systems such as PS3.
52

209. Vector Operations
• The vast majority of the maths mentioned can be
stated as vector and matrix operations.
• Makes the whole architecture very suitable to
execution on Cell SPUs.
‣ Synergistic Processing Units are vector-based
coprocessors in Cell-based systems such as PS3.
• SPUs are typically not used efficiently or fully.
52

210. Vector Operations
• The vast majority of the maths mentioned can be
stated as vector and matrix operations.
• Makes the whole architecture very suitable to
execution on Cell SPUs.
‣ Synergistic Processing Units are vector-based
coprocessors in Cell-based systems such as PS3.
• SPUs are typically not used efficiently or fully.
‣ Effectively “free” processing power.
52

211. Experiments
53

212. Experiments
• Majority of work to date has been conceptual
53

213. Experiments
• Majority of work to date has been conceptual
• Initial prototype has been developed based on a
number of assumptions
53

214. Experiments
• Majority of work to date has been conceptual
• Initial prototype has been developed based on a
number of assumptions
• Work currently ongoing to make sure all of these
assumptions are valid.
53

215. Experiments
• Majority of work to date has been conceptual
• Initial prototype has been developed based on a
number of assumptions
• Work currently ongoing to make sure all of these
assumptions are valid.
• Experiments being run on small problem instances
as translation from SAS+ encoding to internal
architecture rep. currently by hand.
53

216. Results
54

217. Results
• Early tests have shown some promising results
54

218. Results
• Early tests have shown some promising results
‣ Noticeable decrease in time taken to make decisions
over a purely deliberative method.
54

219. Results
• Early tests have shown some promising results
‣ Noticeable decrease in time taken to make decisions
over a purely deliberative method.
- 1-2ms for a 12 decision point execution with 6ms of processing
in advance.
54

220. Results
• Early tests have shown some promising results
‣ Noticeable decrease in time taken to make decisions
over a purely deliberative method.
- 1-2ms for a 12 decision point execution with 6ms of processing
in advance.
‣ Increased robustness to changes detected in the domain
54

221. Results
• Early tests have shown some promising results
‣ Noticeable decrease in time taken to make decisions
over a purely deliberative method.
- 1-2ms for a 12 decision point execution with 6ms of processing
in advance.
‣ Increased robustness to changes detected in the domain
- Rapid discovery of alternative paths through the space
54

222. Results
• Early tests have shown some promising results
‣ Noticeable decrease in time taken to make decisions
over a purely deliberative method.
- 1-2ms for a 12 decision point execution with 6ms of processing
in advance.
‣ Increased robustness to changes detected in the domain
- Rapid discovery of alternative paths through the space
• Much more rigorous testing required.
54

223. Future Work
55

224. Future Work
• A lot of work remains to develop this into a
polished technique that will revolutionise AI.
55

225. Future Work
• A lot of work remains to develop this into a
polished technique that will revolutionise AI.
‣ Further testing on wider range of problems
55

226. Future Work
• A lot of work remains to develop this into a
polished technique that will revolutionise AI.
‣ Further testing on wider range of problems
‣ Full testing of all assumptions made and techniques
chosen without substantiation
55

227. Future Work
• A lot of work remains to develop this into a
polished technique that will revolutionise AI.
‣ Further testing on wider range of problems
‣ Full testing of all assumptions made and techniques
chosen without substantiation
‣ Development into a working system, rather than proof-
of-concept prototype
55

228. Future Work
• A lot of work remains to develop this into a
polished technique that will revolutionise AI.
‣ Further testing on wider range of problems
‣ Full testing of all assumptions made and techniques
chosen without substantiation
‣ Development into a working system, rather than proof-
of-concept prototype
‣ Proof of extensibility of system by addition of additional
Stacks representing other sources of information.
55

229. Questions?