Stefano Nolfi

1
Nolfi S. (2002). Power and limits of reactive
agents. Neurocomputing, 49, 119-145
Also see Scheier C., Pfeifer R., and
Kunyioshi Y. (1998) Embedded Neural Networks
exploiting constraints. Neural Networks 11,
1551-1569.
2
Introduction

• Reactive agents and complex tasks
• No internal representations
• Same reaction to same input
• Sensory-motor coordination
• Pro-active agents

3
Aliasing problem

• Perceptual aliasing
• Active perception as solution

4
Handling overlap

• Type-2 problems -gt Type-1 problems
• Internally recode sensory inputs
• Sensory motor coordination
• Discriminate different objects
• Walls and cilinders
• Big and small cilinders

5
Handling overlap
6
Handling overlap

• Circling behavior
• Geometric Separability Index

Active GSI 0.5 ? 0.9 Passive GSI 0.564722,
0.500556
7
Exploit environment
8
Exploit environment

• All states effected by the aliasing problem

9
Exploit environment

• Produce attractors
• Best action for ambiguous situation
• One action has no function by itself

10
Exploit environment

• Agent has to return to circled zone
• Length walls
• The limit

11
Reactive to Pro-active agents
Environment
Agent
Reactive
i1
o1
o fw(i)
i2
o2
Environment
Pro-active
Agent
o fw(i,s)
i1
o1
s1
i2
o2
12
Pro-active agents

• Introduction of an internal state
• Recurrency
• Neural inertia
• Time delays
• Plastic weights
• Smooth transition to new strategy
• The internal state is not used to encode the
  environment, it is a transformation of past
  experiences and initial state

13
Powers of Pro-active agents

• To take the next step in a bottom-up approach to
  artificial intelligence, the powers of simple
  pro-active agents have to be understood

14
Powers of Pro-active agents