7. A Bottom-up Investigation of Internal Context in an Artificial Agent

7.4. The Environment

The environment is a small artificial stock market. Broadly it is an extension of [14]. There is a small population of artificial traders who are each attempting to predict the next day's stock prices based on data from the recent past. Each trader has a small population of candidate models of the price behaviour in terms of past movement, dividends, comparisons against historical levels etc. Each agent is continually: generating new variations of past models; evaluating them as to how well they would have predicted prices in the recent past; choosing the best of its models; using this model to predict future prices and taking an action to buy or sell dependent on this prediction. It is doing this in parallel to the other agents - essentially each is trying to `out-think' the other traders by predicting the effect of each other's predictions. The price series that results are weakly related to the fundamentals of the stocks (e.g. the dividend to price ratio) but this is `masked' by self-reinforcing patterns of speculation among the traders.

I enhanced this model by substituting a genetic programming algorithm for the classifier system. This enables the traders to develop their internal models in a more open-ended and creative way using a much wider range of strategies, including imitation. I also introduced a round of communication between traders before each session of buying and selling. The details of the traders and their interactions can be found in [6, 7], but they are not critical here. What is important here is that it provides: an environment that is beyond the capacity of agent to completely model [8]; that displays distinct phases to which learning heuristics might apply; and that is tunable in the level and type of learning difficulties it presents. To give an idea of the level of difficulty, figure 6 shows a typical example of the price series that the agent is trying to learn. Although it is fairly unpredictable in detail, there are distinct and recognizable phases of buying and selling indicated by the cyclical nature of price swings and the clear negative correlation between the prices of the two stocks.

The Pragmatic Roots of Context - Bruce Edmonds - 31 MAR 99

[Next] [Previous] [Up] [Top] [Contents]

Generated with CERN WebMaker