Model-Based

[Spyridon Samothrakis]

Hello,

It should be technically feasible to instantiate an "agent within an agent" and do some kind of simulation-based search (e.g. MCTS). Is this "legal" within the rules of the competition?

Thanks,
Spyros

[Jerry Swan]

Hi Spyros,

>Is this "legal" within the rules of the competition?

My personal opinion is that this would be legal, but I can't speak for my co-organizers, so let's see if they have anything to add in the near future.

FWIW, philosophically speaking, my particular interest is in seeing what can be achieved:

a) With as few `hard-coded' assumptions as possible.

b) In a semantics-driven (rather than mechanistic/purely quantative) fashion.

The interest arises precisely because of the tension between a) and b). I am aware that b) is a rather vague statement, but if I knew how to qualify it completely, then I'd probably be both rich and famous...

Claim that MCTS would be in category a) but not b).

Best wishes,

Jerry.

--
You received this message because you are subscribed to the Google Groups "Text Adventure AI Competition" group.
To unsubscribe from this group and stop receiving emails from it, send an email to text-adventure-ai-co...@googlegroups.com.
To post to this group, send email to text-adventure...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/text-adventure-ai-competition/16f9dc99-dd88-4d4a-bf5a-75ce383dde2d%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Tim Atkinson]

Hi Spyros, 

I will have to ask the rest of the team regarding the rules for judging agents that use simulations although we will be limiting computation time during judgement. In the mean time I would say that you can train your agent however you wish to train it, as long as you can separate that trained agent from your simulation-based training when it comes to submitting the entry there should be no problem regardless!

Hope I helped,

Tim

[Spyridon Samothrakis]

Jerry and Tim,

Thank you for your quick responses.

MCTS is called "simulation based search" sometimes because it assumes that the agent has a model inside its "mind". The availability of an internal simulator is a very strong assumption.

Furthermore, no training is required a-priori, unless MCTS is combined with some other method (e.g. RL). If this is OK, then the next question would be "how much computation time is allowed per action?".

Thanks,
Spyros

--

You received this message because you are subscribed to the Google Groups "Text Adventure AI Competition" group.
To unsubscribe from this group and stop receiving emails from it, send an email to text-adventure-ai-co...@googlegroups.com.
To post to this group, send email to text-adventure...@googlegroups.com.

To view this discussion on the web, visit https://groups.google.com/d/msgid/text-adventure-ai-competition/605f09b8-b868-4506-8355-d9274fd633f6%40googlegroups.com.

[Tim Atkinson]

Hi Spyros,

We just had a meeting and decided that simulation was acceptable. You can potentially simulate by playing the game and resetting it with the now-available SoftRestart command and returning to your previous game state by performing the same actions. Obviously how you implement this is up to you. I hope that helps!

Kind regards,

Tim

On Tuesday, 14 June 2016 12:28:04 UTC+1, Spyridon Samothrakis wrote:

Jerry and Tim,

Thank you for your quick responses.

MCTS is called "simulation based search" sometimes because it assumes that the agent has a model inside its "mind". The availability of an internal simulator is a very strong assumption.

Furthermore, no training is required a-priori, unless MCTS is combined with some other method (e.g. RL). If this is OK, then the next question would be "how much computation time is allowed per action?".

Thanks,
Spyros



On 14/06/16 12:21, Tim Atkinson wrote:

Hi Spyros, 

I will have to ask the rest of the team regarding the rules for judging agents that use simulations although we will be limiting computation time during judgement. In the mean time I would say that you can train your agent however you wish to train it, as long as you can separate that trained agent from your simulation-based training when it comes to submitting the entry there should be no problem regardless!

Hope I helped,

Tim

On Tuesday, 14 June 2016 11:43:33 UTC+1, Spyridon Samothrakis wrote:

Hello,

It should be technically feasible to instantiate an "agent within an agent" and do some kind of simulation-based search (e.g. MCTS). Is this "legal" within the rules of the competition?

Thanks,
Spyros

--
You received this message because you are subscribed to the Google Groups "Text Adventure AI Competition" group.

To unsubscribe from this group and stop receiving emails from it, send an email to text-adventure-ai-competition+unsub...@googlegroups.com.