Latest Results

After long months of work on construction of our interface and on bringing the system up to speed, in the past few weeks we have been able to work on fine-tuning our parameters and bringing our agent's scoring percentage to a competitive level.

This was partly done by adding multiple upgrades to our DQN code, which have been implemented on different domains and have shown improvements in results. Some of them are:

• Updates to DQN parameters to reflect the challenges of our domain

• Double DQN

• Success Experience Replay, to enable more focused learning in our large state space

Below are some samples of our results (more to follow soon!):

Empty goal scenario:

Scoring percentages for this scenario have been especially high (even beating some benchmark tests of other systems).

Video of the agent scoring on an empty goal, towards the end of training:

Scoring percentage plot:

1 agent vs. goalie scenario:

Here is a video of the run, as seen sometime around the beginning of the training session (higher quality coming soon):

1 agent vs. goalie and defender scenario:

Scoring percentage plot:

Video coming soon!

Stay tuned for more results and deeper explanations on all scenarios.