Notes completed!

CM3020 Artificial Intelligence/Week 5/Week 5 notes.md

@@ -0,0 +1,46 @@
+# State-of-the-art game players
+Double DQN came to be which led to the Rainbow system.
+
+The DreamerV2 was released
+
+R2D2 also was created
+
+In 2020, Agent57, the best current Atari agent, was released. This made the Atari problem to be considered solved.
+
+AlphaStar aims to solve the StarCraft II game problem. It draws on many areas of AI research, including reinforcement learning, game theory, and evolutionary computation.
+
+OpenAI5 for Dota2 has been created.
+
+The Mario AI competition, which ran from 2009 to 2012, turned into the Platformer AI competition.
+
+# Readings
+## AlphaStar an evolutionary computation perspective
+population-based training (PBT) - keep a population of agents that train against each other
+
+Competitive Co-Evolutionary Algorithms (CCEAs)
+PBT is a memetic algorithm (MA)
+
+It is possible to use self-play to optimize the agent.
+
+Behavior Descriptors (BD)
+
+Quality diversity (QD) algorithms
+
+# Ethics of game playing AI
+> "A completely fair competition can only be achieved against an artificial system that is essentially equivalent ot  aflesh and blood human"
+
+### 6 dimensions of fairness
+1. Perceptual - Do the two competitors use the same input space?
+2. Motoric - Do the two competitors use the same output space?
+3. Historic - Did the two systems spend the same amount of time on training?
+4. Knowledge - Do the two systems have the same access to declarative knowledge about the game?
+5. Compute - Do the two systems have the same computational power?
+6. Common sense - Do the agents have the same knowledge about other things?
+
+### A more moderate view
+> "Through the discussion of two areas (the competitor perceptual and motoric abilities, and the game's extrinsic and intrisic factors) we hope to show that, so far, no fair competition between AI's and humans has occurred in Dota2 and StarCraft II. We thus conclude that we are not done with game."
+
+### Direct impact of superhuman players on the game players
+> The lessons that alphago is teaching use are going to influence how go is played for the next 1000 years"
+
+> The South Korean said he had decided to retire after realising "I'm not at the top even if I become the numberr one"