Add "score" to internal observation.

What exactly NH means by score is a bit context-dependent. The value
u.urexp factors into botl_score() in botl.c (if enabled) as well as
into the computation in really_done() in end.c. In many cases, the
result of botl_score() will be what really_done() reports as the
end-of-game score. In many other cases, it won't: E.g. after dying
(not quitting, escaping, ascending or bug-related exits), score for
acquired gold is reduced by 10%. Also if escaped or ascended, valuable
gems, artefacts etc will give extra score. The computation in end.c
adds to u.urexp, which will reflect the end-of-game score for a few
briefs "steps" at the very end of the game (but not before).

include/nleobs.h

@@ -5,7 +5,7 @@
 #define NLE_MESSAGE_SIZE 256
 #define NLE_BLSTATS_SIZE 25
 #define NLE_PROGRAM_STATE_SIZE 6
-#define NLE_INTERNAL_SIZE 8
+#define NLE_INTERNAL_SIZE 9
 #define NLE_INVENTORY_SIZE 55
 #define NLE_INVENTORY_STR_LENGTH 80
 #define NLE_SCREEN_DESCRIPTION_LENGTH 80

nle/tests/test_profile.py

@@ -2,6 +2,10 @@
 #
 # Copyright (c) Facebook, Inc. and its affiliates.
 
+# Requires
+#   pip install pytest-benchmark
+# to run
+
 import pytest
 
 import numpy as np

win/rl/winrl.cc

@@ -256,6 +256,8 @@ NetHackRL::fill_obs(nle_obs *obs)
         obs->internal[5] = nle_seeds[0]; /* core */
         obs->internal[6] = nle_seeds[1]; /* disp */
         obs->internal[7] = u.uhunger;
+        obs->internal[8] =
+            u.urexp; /* score (careful! check botl_score() and end.c) */
     }
 
     if ((!program_state.something_worth_saving && !program_state.in_moveloop)
@@ -412,7 +414,8 @@ NetHackRL::fill_obs(nle_obs *obs)
         }
     }
     if (obs->screen_descriptions) {
-        memcpy(obs->screen_descriptions, &screen_descriptions_, screen_descriptions_.size());
+        memcpy(obs->screen_descriptions, &screen_descriptions_,
+               screen_descriptions_.size());
     }
 }