Back
to indexElizabeth Sklar, Alan D. Blair, Pablo Funes, and Jordan Pollack
Summary
This paper presents an algorithm for training a population of agents to play competitively against a human at varying levels. The generated population of agents is diverse enough to match an agent with a human of any skill level. Agents learn to play the game of Tron in a restricted online environment. Learning is supervised through a feed-forward neural net which modifies its weights by backprop and Hinton's cross-entropy. The parameters of the neural net include a very low learning rate of 0.0002 and a high momentum of 0.9. These parameters allow the agent to jump quickly to the ideal weights, then stay there for a long duration. Experimental results indicated that this technique produced a sufficiently variable population of skilled agents. The reason I was interested in this paper was for the human interaction, and how the agents learned based on human feedback. Using the methods described below lead to a kind of dead end through the use of a restricted environment. Inability to generalize to a broader population of agents makes this paper rather difficult to build upon.
Methods
The agents learn through a feed-forward neural network, based on feedback from a human (or another agent) trainer. The most successful technique for the agents to learn was related to the ratio of left, right and forward moves to the total set of moves. By this I mean the agent maintained probabilities for each move in the game, and was most successful when it duplicated the same probabilities that the trainer indicated. Winning the game was not the only objective, since creating a diverse population of agents to play humans of different skill levels was. Iterating this probabilistic move recommendation for different skill levels allowed a sufficiently variable population of agents to emerge.
Keywords
human-agent interaction, neural networks, learning
Rating
5
Bibtex Entry
@article = { sklar99,
author = "Elizabeth Sklar and Alan D. Blair and Pablo Funes and Jordan Pollack",
title = "Training Intelligent Agents using Human Internet Data",
journal = "IAT-99",
year = "1999",
url = "http://www.cogs.susx.ac.uk/users/ezequiel/alife-page/alife.html"
}