Update TRAINING.md

TRAINING.md

@@ -18,7 +18,7 @@ To get a sense of the sample efficiency of the standard [PPO algorithm](https://
 
 ![ppo_training](figure/ppo_results.svg)
 
-Out of 17 trials with different initial random seeds, the best one solved the task in 1.274M timesteps, and the median number of timesteps is 2.998M. On a single CPU machine, the wall clock speed to train 1M steps is roughly 1 hour, so we can expect to see the agent learning a reasonable policy after a few hours of training. It is interesting to note that some trials took PPO a long time to learn a reasonable strategy, and it could be due to the fact that we are training a randomly initialized network that knows nothing about Slime Volleyball, against an expert player right at the beginning. It's like an infant learning to play volleyball against an Olympic gold medalist. Here, our agent will likely receive the lowest possible score all the time regardless of any small improvement, making it difficult to learn from constant failure. That PPO still manages to eventually find a good policy is a testament of how good it is. This is an important point we will discuss again later on again, and also one of the reasons self-play can be a useful and powerful approach.
+Out of 17 trials with different initial random seeds, the best one solved the task in 1.274M timesteps, and the median number of timesteps is 2.998M. On a single CPU machine, the wall clock speed to train 1M steps is roughly 1 hour, so we can expect to see the agent learning a reasonable policy after a few hours of training. It is interesting to note that some trials took PPO a long time to learn a reasonable strategy, and it could be due to the fact that we are training a randomly initialized network that knows nothing about Slime Volleyball, against an expert player right at the beginning. It's like an infant learning to play volleyball against an Olympic gold medalist. Here, our agent will likely receive the lowest possible score all the time regardless of any small improvement, making it difficult to learn from constant failure. That PPO still manages to eventually find a good policy is a testament of how good it is. This is an important point that we will discuss again later on, and also one of the reasons self-play can be a useful and powerful approach.
 
 In addition to sample efficiency, we want to know what the best possible performance we can get out of PPO. We ran multi-processor PPO (see [code](https://github.com/hardmaru/slimevolleygym/blob/master/training_scripts/train_ppo_mpi.py)) on a 96-core CPU machine for a while and achieved an average score of 1.377 ± 1.133 over 1000 trials. The highest possible score is 5.0.