Polkadot JAM Slides - Token2049 - By Dr. Gavin Wood
Interactive Artificial Learning in Multi-agent Systems
1. Interactive Artificial Learning in Multi-agent Systems
Yomna M. Hassan, Salman Ahmed, and Jacob W. Crandall
Computing and Information Science Program at the Masdar Institute of Science and Technology, Abu Dhabi, UAE.
Email: {yhassan, sahmed, jcrandall}@masdar.ac.ae
Introduction
Many real-world problems, in which intelligent
Interactive artificial learning process
machines can be useful, require interactions between
multiple intelligent agents. To overcome the challenges
Learning By
Reward
Demonstration Reinforcement
of previously used methods, we are adapting
Interactive artificial learning (IAL) as a learning
methodology in multi-agent systems (MAS). Learning
Step 1
Configure
by demonstration (LbD) and reward reinforcement
Step 2
Plan
Step 3
Step 4
Act
Step 5
Observe
and
Reward
Update
End User
have been studied previously in single agent
environments. We are focusing on MAS.
Learning By Demonstration in Repeated stochastic games
We have performed preliminary investigating the usefulness of LbD in MAS. The simulation have been done on a repeated stochastic
games based which models the iterative prisoner’s dilemma..Results show that LbD helps learning agents learn non-myopic equilibrium
in repeated stochastic games when human demonstrations are
well-informed. On the other hand, when human demonstrations
are less informed, these agents sometimes learn behavior that
produces (less-successful) myopic
behavior.
Human input
initialization
Evaluate fittnes
Check termination
yes
conditiion criteria
terminate
Multi-agent learning algorithms for coordination in smart power grids
no
selection
On-going Research
Human input
In power systems with renewable energy resources, demand response programs can be used
to are encourage more efficient use of energy resources. Intelligent devices can be developed
to help users respond effectively. One method we are considering for these devices is
crossover
interactive evolutionary learning, wherein human input is provided to a genetic algorithm. We
are developing interactive evolutionary algorithms that learn successfully in multi-agent
New population
systems with minimal human input. Basic structure of the algorithm is shown in the figure to
mutation
the left.
Configure
Needs
Demonstration
Task Scheduling in Multi-Vehicle Transportation Systems
( Chernova and Veloso, 2009)
NO
Act
Yes
In general, machine learning algorithms rely heavily on the configuration stage,
Train
wherein the programmer selects relevant features and a distance metric. We are
Needs Help to
Update Policy
investigating the possibility of deriving the distance metric from interactions between
the agent and the user. In particular, we are adapting and extending the CBA
Yes
NO
Update
Policy
Programmer Input
Required
User Input Required
Update Policy
algorithm (Chernova and Veloso, 2009) for an online taxi problem.
Work Autonomsly
Modification Planned