Comparison

But why so many RL algorithms ?

Different tradeoffs

Sample efficiency

Sample efficiency

• How many samples do we need to get a good policy ?

• Off-policy or On-policy ?

  • Off-policy: able to improve the policy without generating new samples from that policy.

  • On-policy: each time the policy is changes, even a little bit, we need to generate new samples.

off/on -policy

But why we would use a less efficient algorithm?

Because sample efficiency is not the only measurement for a RL algorithm, and perhaps the less efficient algorithms are quicker -- wall clock time is not the same as efficiency.

Stability & ease of use

Converge ? Converge to what ? Converge every time ?

Supervised learning is almost always gradient descent, but RL is often not gradient descent. For example, Q-learning is fixed point iteration. And ...

• Policy gradient

  • The only one that actually performs gradient descent (ascent) on the true objective, but also often the least efficient.

• Value function fitting

  • At best, minimizes error of fit ("Bellman error", not the same as expected reward)

  • At worst, doesn't optimize anything, not guaranteed to converge to anything in the nonlinear case.

• Model-based RL

  • Model minimizes error of fit, which will converge.

  • But no guarantee that better model is better policy.

Different assumptions

Stochastic or deterministic/Continuous or discrete/Episodic or infinite horizon

1. Full observability

   • Generally assumed by value function fitting methods

   • Can be mitigated by adding recurrence

2. Episodic learning

   • Often assumed by pure policy gradient methods

   • Assumed by some model-based RL methods

3. Continuity or smoothness

   • Assumed by some continuous value function learning methods

   • Often assumed by some model-based RL methods

Different things are easy or hard in different settings

• Easier to represent the policy

• Easier to represent the model