RL Agent

An RL agent may include one or more of these components:

• Policy: agent’s behaviour function
• Value Function: how good is each state and/or action
• Model: agent’s representation of the environment

Model

It is not required, this is where there is the difference between model-free and model-based RL algorithms.

Categories of RL Agents

• Value-Based $\to$
  • No policy (Implicit, ex: $ϵ$-greedy), there is Value Function
• Policy-Based $\to$ Policy Gradient Methods
  • Policy, and no Value Function
• Actor Critic

  • Stores both policy and value function

  • Critic: Updates action-value function parameters $w$

  • Actor: Updates policy parameters $\theta$, in direction suggested by critic

  • Actor-critic algorithms follow an approximate policy gradient

Model-Free → Policy and/or value function, no model Model-Based → Policy and/or value function, with model

Policy Gradient Methods vs. Value-Based Methods

Why would you choose a value-based method over a policy-based method?

Policy optimization:

• More stable
• Less sample-efficient

Value-based methods:

• more sample-efficient (because off-policy, old samples can still be reused, there’s a Replay buffer)

The value function might be quite complicated. But the policy might be very simple to represent. There are cases where a policy might be more compact.

See Policy Gradient Methods for advantages and disadvantages.

The problem value-based methods

In high dimensions, there can be many maximums for the Q-function that be “hacked”, resulting in exploits