Thank you for taking the time to respond, I appreciate it.
Well I guess my question regarding the expensiveness comes down to wondering about the sample efficiency, i.e. are there not many games that share large similar state trajectories that can be re-used? Are you using any off-policy corrections, e.g. IMPALA style?
Or is that just a source off noise that is too difficult to deal with and/or the state space is so large and diverse that that many samples are really needed? Maybe my intuition is just way off, it just feels like a very very large sample size.
Reminds me slightly of the first version of the non-hierarchical TensorFlow device placement work which needed a fair bit of samples, and a large sample efficiency improvement in the subsequent hierarchical placer. So I recognise there is large value in knowing the limits of a non-hierarchical model now and subsequent models should rapidly improve sample efficiency by doing similar task decomposition?
Well I guess my question regarding the expensiveness comes down to wondering about the sample efficiency, i.e. are there not many games that share large similar state trajectories that can be re-used? Are you using any off-policy corrections, e.g. IMPALA style?
Or is that just a source off noise that is too difficult to deal with and/or the state space is so large and diverse that that many samples are really needed? Maybe my intuition is just way off, it just feels like a very very large sample size.
Reminds me slightly of the first version of the non-hierarchical TensorFlow device placement work which needed a fair bit of samples, and a large sample efficiency improvement in the subsequent hierarchical placer. So I recognise there is large value in knowing the limits of a non-hierarchical model now and subsequent models should rapidly improve sample efficiency by doing similar task decomposition?