By%20Philip%20Sterne

1
Sharing Learned Information
An examination of reinforcement learning using
neural networks in a multi-agent foraging task

• 6. Learning Algorithm
• The Neural Net is initialized randomly.
• In each turn
• Obtain Sight Hearing inputs
• Obtain outputs from Neural network
• Perform Collision Detection and calculate
  rewards/penalties
• If the reward is
• Positive Train the network to output a similar
  but larger output.
• Negative Train the network to output an output
  that is orthogonal to the previous output.

1. Statement of intent We are interested in
finding efficient multi-agent foraging strategies
automatically. To achieve this we define which
actions are good, and which are bad. We then let
the agents explore the space, and receive
information about their actions. We use a neural
network to then learn the best actions in any
given circumstance.
By Philip Sterne Email g99s0115_at_campus.ru.ac.za O
r visit www.cs.ru.ac.za\research\g99s0115\ Superv
ised by Prof. S. Bangay
7. Extensions I also intend on exploring
temporal difference learning, where a large
positive, or negative reward affects the training
of the preceding actions.
Implementation This work has been done in Matlab,
making use of the neural networks toolbox.