2-Step Follow the Regularized Leader

We have seen 3 online learning algorithms: 1-step OMD, 2-step OMD and FTRL. Usually FTRL is written and studied as a single step update. However, it is possible to define and analyze a 2-step version of FTRL as well.

The FTRL update is written as:

\[\begin{align*} x_1 &= \arg\min_{x \in \Delta_n} F(x)\\ x_{t+1} &= \arg\min_{x \in \Delta_n} \left[\eta_t \sum_{s=1}^t l_s^\top x + F(x)\right] \end{align*}\]

Using the Mixed Bregman \(\text{Breg}_{a,b}(x\|y) = \frac{F(x)}{a} - \frac{F(y)}{b} - \frac{\nabla F(y)}{b}^\top(x-y)\), we can write FTRL in a way similar to 1-step OMD.

\[\begin{align*} x_1 &= \arg\min_{x \in \Delta_n} F(x)\\ x_{t+1} &= \arg\min_{x \in \Delta_n} \left[ l_t^\top x + \text{Breg}_{\eta_t,\eta_{t-1}}(x\|x_t)\right] \end{align*}\]

So, we can write a 2-step FTRL similar to 2-step OMD:

\[\begin{align*} x_1 &= \arg\min_{x \in \Delta_n} F(x)\\ \tilde{x}_{t+1} &= \arg\min_{x \in \mathbb{R}^n} \left[ l_t^\top x + \text{Breg}_{\eta_t,\eta_{t-1}}(x\|x_t)\right]\\ x_{t+1} &= \arg\min_{x \in \Delta_n} \text{Breg}(x\|\tilde{x}_{t+1}) \end{align*}\]

Here is a summary of the four algorithms:

Algorithm 1-step 2-step
OMD $$x_{t+1} = \arg\min_{x \in \Delta_n} \left[ l_t^\top x + \text{Breg}_{\eta_t,\eta_{t}}(x\|x_t)\right]$$ $$\tilde{x}_{t+1} = \arg\min_{x \in \mathbb{R}^n} \left[ l_t^\top x + \text{Breg}_{\eta_t,\eta_{t}}(x\|x_t)\right]$$$$x_{t+1} = \arg\min_{x \in \Delta_n} \text{Breg}(x\|\tilde{x}_{t+1})$$
FTRL $$x_{t+1} = \arg\min_{x \in \Delta_n} \left[ l_t^\top x + \text{Breg}_{\eta_t,\eta_{t-1}}(x\|x_t)\right]$$ $$\tilde{x}_{t+1} = \arg\min_{x \in \mathbb{R}^n} \left[ l_t^\top x + \text{Breg}_{\eta_t,\eta_{t-1}}(x\|x_t)\right]$$$$x_{t+1} = \arg\min_{x \in \Delta_n} \text{Breg}(x\|\tilde{x}_{t+1})$$

In terms of potential functions, they can be written as:

Algorithm 1-step 2-step
OMD $$\theta_t = \theta_{t-1} - \eta_t l_t$$$$x_{t+1} = \psi(\theta_t + \lambda(\theta_t))$$ $$\theta_t = \theta_{t-1} - \eta_t l_t$$$$\tilde{x}_{t+1} = \psi(\theta_t + \lambda(\theta_{t-1}))$$$$x_{t+1} = \psi(\theta_t + \lambda(\theta_{t}))$$
FTRL $$\theta_t = \frac{\eta_t}{\eta_{t-1}}\theta_{t-1} - \eta_t l_t$$$$x_{t+1} = \psi(\theta_t + \lambda(\theta_t))$$ $$\theta_t = \frac{\eta_t}{\eta_{t-1}}\theta_{t-1} - \eta_t l_t$$$$\tilde{x}_{t+1} = \psi\left(\theta_t + \frac{\eta_t}{\eta_{t-1}}\lambda(\theta_{t-1})\right)$$$$x_{t+1} = \psi(\theta_t + \lambda(\theta_{t}))$$

Regret

In previous posts, we have analyzed 1-step OMD, 2-step OMD and 1-step FTRL. Here, we analyze 2-step FTRL

For any \(x \in \Delta_n\), we have:

\[\begin{align*} l_t^\top(x_t-x) &= l_t^\top(\tilde{x}_{t+1}-x) + l_t^\top(x_t-\tilde{x}_{t+1})\\ &= \left(\frac{\theta_{t-1}}{\eta_{t-1}}-\frac{\theta_t}{\eta_t}\right)^\top(\tilde{x}_{t+1}-x) + l_t^\top(x_t-\tilde{x}_{t+1})\\ &= \left(\frac{\nabla F(x_t)-\lambda(\theta_{t-1})}{\eta_{t-1}} - \frac{\nabla F(\tilde{x}_{t+1})-\frac{\eta_t}{\eta_{t-1}}\lambda(\theta_{t-1})}{\eta_t}\right)^\top(\tilde{x}_{t+1}-x) + l_t^\top(x_t-\tilde{x}_{t+1})\\ &= \left(\frac{\nabla F(x_t)}{\eta_{t-1}} - \frac{\nabla F(\tilde{x}_{t+1})}{\eta_t}\right)^\top(\tilde{x}_{t+1}-x) + l_t^\top(x_t-\tilde{x}_{t+1}) \end{align*}\]

Using the Mixed Bregman, we can write the first term as:

The first term can be written as:

\[\begin{align*} \left(\frac{\nabla F(x_t)}{\eta_{t-1}} - \frac{\nabla F(\tilde{x}_{t+1})}{\eta_t}\right)^\top(\tilde{x}_{t+1}-x) &= \text{Breg}_{\alpha,\eta_{t-1}}(x\|x_t) - \text{Breg}_{\alpha,\eta_{t}}(x\|\tilde{x}_{t+1})- \text{Breg}_{\eta_t,\eta_{t-1}}(\tilde{x}_{t+1}\|x_t)\\ &\leq \text{Breg}_{\alpha,\eta_{t-1}}(x\|x_t) - \text{Breg}_{\alpha,\eta_{t}}(x\|x_{t+1})- \text{Breg}_{\eta_t,\eta_{t-1}}(\tilde{x}_{t+1}\|x_t)\\ \end{align*}\]

\(\alpha\) could be any non-zero number in the above expression.

Note that \(\text{Breg}_{\alpha,\eta_{t}}(x\|x_{t+1}) \leq \text{Breg}_{\alpha,\eta_{t}}(x\|\tilde{x}_{t+1})\) for all \(x \in \Delta_n\).

\[\begin{align*} \eta_t(\text{Breg}_{\alpha,\eta_{t}}(x\|\tilde{x}_{t+1}) - \text{Breg}_{\alpha,\eta_{t}}(x\|x_{t+1})) &= F(x_{t+1}) - F(\tilde{x}_{t+1}) - \nabla F(\tilde{x}_{t+1})^\top (x-\tilde{x}_{t+1}) + \nabla F(x_{t+1})^\top (x-x_{t+1}) \\ &= \text{Breg}(x_{t+1}\| \tilde{x}_{t+1}) + ( \nabla F(x_{t+1}) - \nabla F(\tilde{x}_{t+1}))^\top (x-x_{t+1}) \\ &\geq ( \nabla F(x_{t+1}) - \nabla F(\tilde{x}_{t+1}))^\top (x-x_{t+1})\\ &= (\theta_t + \lambda(\theta_t) - \theta_t - \frac{\eta_t}{\eta_{t-1}}\lambda(\theta_{t-1}))^\top (x-x_{t+1})=0 \end{align*}\]

Taking summation over \(t\), we have:

\[\begin{align*} \sum_{t=1}^T l_t^\top(x_t-x) &= \sum_{t=1}^T \left[\text{Breg}_{\alpha,\eta_{t-1}}(x\|x_t) - \text{Breg}_{\alpha,\eta_{t}}(x\|x_{t+1})\right] + \sum_{t=1}^T \left[ l_t^\top(x_t-\tilde{x}_{t+1}) - \text{Breg}_{\eta_t,\eta_{t-1}}(\tilde{x}_{t+1}\|x_t)\right]\\ &=\text{Breg}_{\alpha,\eta_{0}}(x\|x_1) - \text{Breg}_{\alpha,\eta_{T}}(x\|x_{T+1}) + \sum_{t=1}^T \left[ l_t^\top(x_t-\tilde{x}_{t+1}) - \text{Breg}_{\eta_t,\eta_{t-1}}(\tilde{x}_{t+1}\|x_t)\right] \end{align*}\]

Let \(\alpha = \eta_0 = \eta_T\). The Mixed Bregmans in the first term will become normal Bregmans.

\[\begin{align*} \sum_{t=1}^T l_t^\top(x_t-x) &=\frac{1}{\eta_T}\text{Breg}(x\|x_1) - \text{Breg}(x\|x_{T+1}) + \sum_{t=1}^T \left[ l_t^\top(x_t-x_{t+1}) - \text{Breg}_{\eta_t,\eta_{t-1}}(x_{t+1}\|x_t)\right]\\ &\leq \frac{1}{\eta_T}\text{Breg}(x\|x_1) + \sum_{t=1}^T \left[ l_t^\top(x_t-\tilde{x}_{t+1}) - \text{Breg}_{\eta_t,\eta_{t-1}}(\tilde{x}_{t+1}\|x_t)\right]\\ \end{align*}\]

Inorder to get specific results, we proceed in the same fashion as 1-step FTRL. For the three cases, we can get similar results with \(x_{t+1}\) replaced by \(\tilde{x}_{t+1}\) in the second summation.

Here is a summary of the regret inequalitis:

Algorithm Regret
1-step OMD $$\sum_{t=1}^T \frac{1}{\eta_t}\left[ \text{Breg}(x\|x_{t}) - \text{Breg}(x\|x_{t+1}) \right] + \sum_{t=1}^T \left[ l_t^\top(x_t-x_{t+1}) - \frac{1}{\eta_t}\text{Breg}(x_{t+1}\|x_{t}) \right]$$
2-step OMD $$\sum_{t=1}^T \frac{1}{\eta_t}\left[ \text{Breg}(x\|x_{t}) - \text{Breg}(x\|x_{t+1}) \right] + \sum_{t=1}^T \left[ l_t^\top(x_t-\tilde{x}_{t+1}) - \frac{1}{\eta_t}\text{Breg}(\tilde{x}_{t+1}\|x_{t}) \right]$$
1-step FTRL $$\frac{1}{\eta_T}\text{Breg}(x\|x_1) + \sum_{t=1}^T \left[ l_t^\top(x_t-x_{t+1}) - \text{Breg}_{\eta_t,\eta_{t-1}}(x_{t+1}\|x_t)\right]$$
2-step FTRL $$\frac{1}{\eta_T}\text{Breg}(x\|x_1) + \sum_{t=1}^T \left[ l_t^\top(x_t-\tilde{x}_{t+1}) - \text{Breg}_{\eta_t,\eta_{t-1}}(\tilde{x}_{t+1}\|x_t)\right]$$

Parts of this post appear in my paper : Putta, S. R., & Agrawal, S. (2021). Scale Free Adversarial Multi Armed Bandits. arXiv preprint arXiv:2106.04700.

Written on May 20, 2021