flâneur

What is AIXI?

jan.leike.name · 1,927 words · saved by 2 readers

This is a short introduction to a mathematical theory of artificial general intelligence that is meant to be accessible without much prior knowledge in mathematics or machine learning. In reinforcement learning an agent (e.g. a robot) interacts with an environment (the space the agent lives in). For the time steps t=1,2,… 𝑡 = 1 , 2 , … , the agent takes an action a t 𝑎 𝑡 (moving its actuators/limbs) and receives a percept e t =( o t , r t ) 𝑒 𝑡 = ( 𝑜 𝑡 , 𝑟 𝑡 ) consisting of an observation o t 𝑜 𝑡 (sensory input like a camera image) and a numeric reward r t ∈R 𝑟 𝑡 ∈ 𝑅 . You can think of rewards as treats for the agent; the larger the number, the better the treat. We usually assume that rewards are bounded between 0 and 1 (mostly for technical reasons). The goal is to maximize total rewards ∑ m t=1 r t ∑ 𝑡 = 1 𝑚 𝑟 𝑡 until some pre-specified time limit m 𝑚 is reached. It is important to emphasize that rewards are all that our agent cares about. Sometimes we

What is AIXI? What is AIXI? — An Introduction to General Reinforcement Learning Jan Leike, December 2015 This is a short introduction to a mathematical theory of artificial general intelligence that is meant to be accessible without much prior knowledge in mathematics or machine learning. Reinforcement Learning In reinforcement learning an agent (e.g. a robot) interacts with an environment (the space the agent lives in). For the time steps \(t = 1, 2, \ldots\), the agent takes an action \(a_t\) (moving its actuators/limbs) and receives a percept \(e_t = (o_t, r_t)\) consisting of an observatio

saved by

related reading