From 65c1550cfaee89c980a7b9f722e8925363507834 Mon Sep 17 00:00:00 2001
From: andrewcoh <54679309+andrewcoh@users.noreply.github.com>
Date: Wed, 17 Mar 2021 17:46:15 -0400
Subject: [PATCH] R15 fix elo (#5151)

* add group done to ELO computation

* add not interrupted
---
 ml-agents/mlagents/trainers/ghost/trainer.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/ml-agents/mlagents/trainers/ghost/trainer.py b/ml-agents/mlagents/trainers/ghost/trainer.py
index cd08ee24f0..2449734b75 100644
--- a/ml-agents/mlagents/trainers/ghost/trainer.py
+++ b/ml-agents/mlagents/trainers/ghost/trainer.py
@@ -190,7 +190,11 @@ def _process_trajectory(self, trajectory: Trajectory) -> None:
         i.e. in asymmetric games. We assume the last reward determines the winner.
         :param trajectory: Trajectory.
         """
-        if trajectory.done_reached:
+        if (
+            trajectory.done_reached
+            and trajectory.all_group_dones_reached
+            and not trajectory.interrupted
+        ):
             # Assumption is that final reward is >0/0/<0 for win/draw/loss
             final_reward = (
                 trajectory.steps[-1].reward + trajectory.steps[-1].group_reward