Progress is going well on the new system, and it appears to be performing well. I just today finished up the mutations and will be moving onto modularisation, which is likely to cause problems. Still, once I have that resolved, I’ll need to move onto other directions.
Sure, there’s all the experiments I can run, like observing the effect of population size/elite percentage, and there’s getting Ms. PacMan and StarCraft running, but those will likely only take up 2010, or less. So the problem now lies in what I can do in the future.
One area is to try to remove the ‘bonuses’ the agent is given for it’s learning (valid actions, initial pre-goal, known goal state). Each of these can safely be removed, but it makes things much more difficult for the agent, as in all cases it has to learn the dynamics of the environment by itself, probably with an alternative learner (TG or RIBS, etc).
Another area is to try and ‘commercialise’ the agent. In the sense that WEKA is a machine learning suite, the agent and environment could be loaded from GUI form and run, allowing the user to specify the necessary information (and be allowed the option of leaving out some things, as mentioned above). This could also tie into teacher/student learning, allowing the user to interrupt the agent and teach it what to do at a given point.
If I was ever to get stuck, I’m sure I could just look back at my posts from earlier in the research to attempt to solve some of the earlier dreams. It would be really cool if I could get a robot to play around with.