Algorithm 1.

Action-Centered Thompson Sampling

1:	Set B = I, θ̂ = 0, b̂ = 0, choose [πmin, πmax].
2:	for t = 1, 2,… do
3:	Observe current context s̄t and form st,a for each a ∈ {1,…, N}.
4:	Randomly generate θ′ ~ 𝒩(θ̂, v2B−1).
5:	Let
	$${\overline{a}}_t=arg\underset{a\in \{1,\dots ,N\}}{max}{s}_{t,a}^T{\theta }^{\prime }.$$
6:	Compute probability πt of taking a nonzero action according to (6).
7:	Play action at = āt with probability πt, else play at = 0.
8:	Observe reward rt(at) and update θ̂
	$$B=B+{\pi }_t(1-{\pi }_t)s_{t,{\overline{a}}_t}{s}_{t,{\overline{a}}_t}^T,\widehat{b}=\widehat{b}+s_{t,{\overline{a}}_t}(I(a_t>0)-{\pi }_t)r_t(a_t),\widehat{\theta }=B^{-1}\widehat{b}.$$
	end for