report.txt

Observation:
Running the testrunner multiple times resulted in two exact runs.
By observation, the behavior of the agent seemed identical between the runs.
It followed the same path(by performing the same actions) and achieved the same scores
for the two episodes. Episode 1 in run 1 matched episode 1 in run 2 and the same for the second episode.
I'm suspecting the first versions of the network/algorithm may suffer from high variance.
The agent hasn't learned how to find bananas it's learned how to follow good paths that lead to bananas.
A few potential causes and fixes;
Perhaps the initial decay of epsilon is too high and it's minimum value is too low 
so that the agent performs too little initial and continued exploration.
Increase network size and depth to "hold more data" and allow more advanced strategies to be developed.
Decrease learning rate.
This can perhaps also be addressed by bootstrapping the targets with multiple steps
as in A3C.

The agent may also get stuck between two bananas alternating between going left and right.