 |
Email:
Homepage: http://wwwhome.ewi.utwente.nl/~vrielink
|
|
| Beschrijving: |
In modern computer games, the application of finite-state machines for artificial intelligence is no longer sufficient. A possible alternative to finite-state machines is to model the environment as a (PO)MDP and use methods that try to maximize rewards over a period of time. With these methods, the agent tries to learn a strategy (called a policy) that maximizes the reward he gains from his actions during a certain period of his lifetime. The solutions found by these methods are often near optimal and thus promising.
A problem with these (PO)MDP methods is that they require vast amounts of calculations: for each step further into the future the agent must take each possible action, resulting state and possible observation into account. Executing such methods often requires linear programming, which is costly in many dimensions.
Because computer games often have an enormous amount of possible world states and are usually partially observable, the previous problem become very apparent for such games. A possible solution to keep games more tractable is to avoid the creation of a policy that spans the entire game world. The assignment tries to achieve this solution by dividing the game world into so called ‘game situations’ and to only create a policy for the game situation the agent is currently in. A policy of such a game situation can be seen as a partial policy of the complete policy that spans the entire world. The assignment tries to find an answer for the following questions:
1. Is it possible to divide an MDP environment into game situations, for which near optimal policy’s can be found?
2. How can these partial policy’s be used by the agent?
3. What advantages and disadvantages does this approach have with respect to methods that create a single policy over the entire game world.
4. How does an agent using this method fare against a finite-state agent in a game of "Capture the flag"
|
|
| Beschrijving: |
De meest gebruikelijke manieren om in de civiele wereld gedistribueerde simulaties te doen is door middel van een peer-to-peer of client-server architectuur. Voor militaire simulatoren (vb: JSF simulator) wordt gebruik gemaakt van IEEE 1516, High Level Architecture (HLA). Dit is een standaard waarmee een stelsel van simulaties beschreven kan worden.
De opdracht bestaat uit:
1. het maken van een overzicht wat HLA is
2. het vergelijken van standaard civiele simulatiemethoden tov HLA
3. het eventueel implementeren van een prototype dmv. HLA
|
|
|
