JESS: Using JESS for representing game states

classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

JESS: Using JESS for representing game states

Sam Sarjant
Hello,
  I have been using JESS for some time as a rule-system for representing states in a game (relational reinforcement learning), such that the state of the game can be represented in terms of objects. An agent uses these states and a decision making process (in this case defqueries to resolve rules in conditions -> action form). Because games are very dynamic environments, the state is constantly changing. I am currently handling this by resetting the state and reasserting the facts of the state (then running to generate any further facts that rules generate).

My question is: is JESS the best rule-system to use for this task? I know JESS has the property of only generating facts once, but if the state is constantly being reset, perhaps some other system would be more effective. JESS certainly gets the job done, and I haven't really tested anything else, but I am concerned with the speed of execution, but perhaps this is simply due to the fact that I am using this relational representation.

--
- Thanks, Sam Sarjant
Reply | Threaded
Open this post in threaded view
|

RE: JESS: Using JESS for representing game states

Friedman-Hill, Ernest
It is true that the Rete algorithm (on which Jess is based) is built on the assumption that only a small fraction (usually quoted as 5-10%) of the knowledge base will change on each evaluation cycle. Populating the network from scratch is expensive and constantly resetting it does degrade performance quite seriously.
 
On the other hand, if the whole knowledge base must be constantly reevaluated, than no other algorithm will necessarily be any better. Rather than searching for other tools, you might put some work into figuring out how to preserve those parts of the KB that don't change as frequently, rather than constantly resetting. Presumably there are stationary objects and information about ownership and properties that don't change often.


From: [hidden email] [mailto:[hidden email]] On Behalf Of Sam Sarjant
Sent: Sunday, October 09, 2011 5:49 PM
To: jess-users
Subject: JESS: Using JESS for representing game states

Hello,
  I have been using JESS for some time as a rule-system for representing states in a game (relational reinforcement learning), such that the state of the game can be represented in terms of objects. An agent uses these states and a decision making process (in this case defqueries to resolve rules in conditions -> action form). Because games are very dynamic environments, the state is constantly changing. I am currently handling this by resetting the state and reasserting the facts of the state (then running to generate any further facts that rules generate).

My question is: is JESS the best rule-system to use for this task? I know JESS has the property of only generating facts once, but if the state is constantly being reset, perhaps some other system would be more effective. JESS certainly gets the job done, and I haven't really tested anything else, but I am concerned with the speed of execution, but perhaps this is simply due to the fact that I am using this relational representation.

--
- Thanks, Sam Sarjant
Reply | Threaded
Open this post in threaded view
|

RE: JESS: Using JESS for representing game states

John Everett-2

In this situation, you might want to investigate the use of the logical conditional element to link dependent  objects in your game state. Taking the example from the Jess manual on page 42, you assert that (water-flowing) is dependent on (faucet-open). Then if you assert or retract (faucet-open) Jess will automatically update the state to reflect the ramifications for that (i.e., whether or not water is flowing). I think you could use this mechanism to create and then incrementally modify your game state rather than reasserting from scratch all the time.

 

From: Friedman-Hill, Ernest [mailto:[hidden email]]
Sent: Monday, October 10, 2011 10:15 AM
To: jess-users
Subject: RE: JESS: Using JESS for representing game states

 

It is true that the Rete algorithm (on which Jess is based) is built on the assumption that only a small fraction (usually quoted as 5-10%) of the knowledge base will change on each evaluation cycle. Populating the network from scratch is expensive and constantly resetting it does degrade performance quite seriously.

 

On the other hand, if the whole knowledge base must be constantly reevaluated, than no other algorithm will necessarily be any better. Rather than searching for other tools, you might put some work into figuring out how to preserve those parts of the KB that don't change as frequently, rather than constantly resetting. Presumably there are stationary objects and information about ownership and properties that don't change often.

 


From: [hidden email] [hidden email] On Behalf Of Sam Sarjant
Sent: Sunday, October 09, 2011 5:49 PM
To: jess-users
Subject: JESS: Using JESS for representing game states

Hello,

  I have been using JESS for some time as a rule-system for representing states in a game (relational reinforcement learning), such that the state of the game can be represented in terms of objects. An agent uses these states and a decision making process (in this case defqueries to resolve rules in conditions -> action form). Because games are very dynamic environments, the state is constantly changing. I am currently handling this by resetting the state and reasserting the facts of the state (then running to generate any further facts that rules generate).

 

My question is: is JESS the best rule-system to use for this task? I know JESS has the property of only generating facts once, but if the state is constantly being reset, perhaps some other system would be more effective. JESS certainly gets the job done, and I haven't really tested anything else, but I am concerned with the speed of execution, but perhaps this is simply due to the fact that I am using this relational representation.

 

--
- Thanks, Sam Sarjant