Okay, so this may be a really stupid question, but would it be possible to use botworld to test decision theories experimentally? Like pitting a bot running TDT against one running CDT in a bunch of partially-randomized scenarios, and seeing who scores higher?
For a given situation, you might be able to work out “what would a TDT agent do”, and you might be able to program your bot do that. (These are different challenges, because a TDT agent might analyse its opponent’s source code, which is a difficult programming problem.) You might be able to do the same for CDT, and pit the two bots against each other. I guess you wouldn’t learn very much from the results, because if you can program the TDT and CDT actions for this situation, it’s probably simple enough to just analyze it and work out what happens.
Implementing either of these algorithms in general is beyond our current abilities.
Okay, so this may be a really stupid question, but would it be possible to use botworld to test decision theories experimentally? Like pitting a bot running TDT against one running CDT in a bunch of partially-randomized scenarios, and seeing who scores higher?
For a given situation, you might be able to work out “what would a TDT agent do”, and you might be able to program your bot do that. (These are different challenges, because a TDT agent might analyse its opponent’s source code, which is a difficult programming problem.) You might be able to do the same for CDT, and pit the two bots against each other. I guess you wouldn’t learn very much from the results, because if you can program the TDT and CDT actions for this situation, it’s probably simple enough to just analyze it and work out what happens.
Implementing either of these algorithms in general is beyond our current abilities.